Modular Forms, a Computational Approach (Graduate Studies in Mathematics) (Graduate Studies in Mathematics, 79) 0821839608, 9780821839607

Citation preview

Modular Forms: A Computational Approach William A. Stein (with an appendix by Paul E. Gunnells) Department of Mathematics, University of Washington E-mail address: [email protected] Department of Mathematics and Statistics, University of Massachusetts E-mail address: [email protected]

1991 Mathematics Subject Classification. Primary 11; Secondary 11-04 Key words and phrases. abelian varieties, cohomology of arithmetic groups, computation, elliptic curves, Hecke operators, modular curves, modular forms, modular symbols, Manin symbols, number theory Abstract. This is a textbook about algorithms for computing with modular forms. It is nontraditional in that the primary focus is not on underlying theory; instead, it answers the question “how do you explicitly compute spaces of modular forms?”

v

To my grandmother, Annette Maurer.

Contents

Preface Chapter 1.

xi Modular Forms

1

§1.1. Basic Definitions

1

§1.3. Modular Forms of Any Level

4

§1.2. Modular Forms of Level 1

3

§1.4. Remarks on Congruence Subgroups

7

§1.5. Applications of Modular Forms §1.6. Exercises Chapter 2.

9 11

Modular Forms of Level 1

13

§2.1. Examples of Modular Forms of Level 1

13

§2.3. The Miller Basis

20

§2.5. Computing Hecke Operators

26

§2.7. Fast Computation of Bernoulli Numbers

29

§2.2. Structure Theorem for Level 1 Modular Forms

17

§2.4. Hecke Operators

22

§2.6. Fast Computation of Fourier Coefficients

29

§2.8. Exercises

33

Chapter 3.

Modular Forms of Weight 2

35

§3.1. Hecke Operators

36

§3.3. Computing with Modular Symbols

41

§3.2. Modular Symbols

39

vii

viii

Contents

§3.4. Hecke Operators

47

§3.6. Computing a Basis for S2 (Γ0 (N ))

53

§3.8. Exercises

60

§3.5. Computing the Boundary Map

§3.7. Computing S2 (Γ0 (N )) Using Eigenvectors

Chapter 4.

Dirichlet Characters

51 58

63

§4.1. The Definition

64

§4.3. Evaluation of Dirichlet Characters

67

§4.2. Representing Dirichlet Characters

64

§4.4. Conductors of Dirichlet Characters

70

§4.5. The Kronecker Symbol

72

§4.7. Alternative Representations of Characters

77

§4.6. Restriction, Extension, and Galois Orbits §4.8. Dirichlet Characters in SAGE

§4.9. Exercises Chapter 5.

Eisenstein Series and Bernoulli Numbers

75 78 81 83

§5.1. The Eisenstein Subspace

83

§5.3. Explicit Basis for the Eisenstein Subspace

88

§5.2. Generalized Bernoulli Numbers

83

§5.4. Exercises

90

Chapter 6.

Dimension Formulas

91

§6.1. Modular Forms for Γ0 (N )

92

§6.3. Modular Forms with Character

98

§6.2. Modular Forms for Γ1 (N ) §6.4. Exercises

Chapter 7.

Linear Algebra

§7.1. Echelon Forms of Matrices

95 102 103 103

§7.2. Rational Reconstruction

105

§7.4. Echelon Forms via Matrix Multiplication

110

§7.3. Echelon Forms over Q

107

§7.5. Decomposing Spaces under the Action of Matrix

114

§7.6. Exercises Chapter 8.

General Modular Symbols

119 121

ix

Contents

§8.1. Modular Symbols

122

§8.3. Hecke Operators

128

§8.5. Pairing Modular Symbols and Modular Forms

137

§8.2. Manin Symbols

124

§8.4. Cuspidal Modular Symbols

133

§8.6. Degeneracy Maps

142

§8.8. Explicit Examples

147

§8.10.

Applications

155

Exercises

156

§8.7. Explicitly Computing Mk (Γ0 (N ))

§8.9. Refined Algorithm for the Presentation §8.11.

Chapter 9.

Computing with Newforms

§9.1. Dirichlet Character Decomposition

144 154

159 159

§9.2. Atkin-Lehner-Li Theory

161

§9.4. Congruences between Newforms

170

§9.3. Computing Cusp Forms §9.5. Exercises

Chapter 10.

Computing Periods

165 176 177

§10.1.

The Period Map

178

§10.2.

Abelian Varieties Attached to Newforms

178

§10.3.

Extended Modular Symbols

179

§10.4.

Approximating Period Integrals

180

§10.5.

Speeding Convergence Using Atkin-Lehner

183

§10.6.

Computing the Period Mapping

185

§10.7.

All Elliptic Curves of Given Conductor

187

§10.8.

Exercises

190

Chapter 11.

Solutions to Selected Exercises

191

§11.1.

Chapter 1

191

§11.2.

Chapter 2

193

§11.3.

Chapter 3

194

§11.4.

Chapter 4

196

§11.5.

Chapter 5

197

§11.6.

Chapter 6

197

§11.7.

Chapter 7

198

x

Contents

§11.8.

§11.9.

§11.10.

Chapter 8

199

Chapter 9

201

Chapter 10

Appendix A.

Computing in Higher Rank

201 203

§A.1. Introduction

203

§A.3. Combinatorial Models for Group Cohomology

213

§A.2. Automorphic Forms and Arithmetic Groups

205

§A.4. Hecke Operators and Modular Symbols

225

§A.5. Other Cohomology Groups

§A.6. Complements and Open Problems

232 244

Bibliography

253

Index

265

Preface

This is a graduate-level textbook about algorithms for computing with modular forms. It is nontraditional in that the primary focus is not on underlying theory; instead, it answers the question “how do you use a computer to explicitly compute spaces of modular forms?” This book emerged from notes for a course the author taught at Harvard University in 2004, a course at UC San Diego in 2005, and a course at the University of Washington in 2006. The author has spent years trying to find good practical ways to compute with classical modular forms for congruence subgroups of SL2 (Z) and has implemented most of these algorithms several times, first in C++ [Ste99b], then in MAGMA [BCP97], and as part of the free open source computer algebra system SAGE (see [Ste06]). Much of this work has involved turning formulas and constructions buried in obscure research papers into precise computational recipes then testing these and eliminating inaccuracies. The author is aware of no other textbooks on computing with modular forms, the closest work being Cremona’s book [Cre97a], which is about computing with elliptic curves, and Cohen’s book [Coh93] about algebraic number theory. In this book we focus on how to compute in practice the spaces Mk (N, ε) of modular forms, where k ≥ 2 is an integer and ε is a Dirichlet character of modulus N (the appendix treats modular forms for higher rank groups). We spend the most effort explaining the general algorithms that appear so far to be the best (in practice!) for such computations. We will not discuss in any detail computing with quaternion algebras, half-integral weight forms, weight 1 forms, forms for noncongruence subgroups or groups other

xi

xii

Preface

than GL2 , Hilbert and Siegel modular forms, trace formulas, p-adic modular forms, and modular abelian varieties, all of which are topics for additional books. We also rarely analyze the complexity of the algorithms, but instead settle for occasional remarks about their practical efficiency. For most of this book we assume the reader has some prior exposure to modular forms (e.g., [DS05]), though we recall many of the basic definitions. We cite standard books for proofs of the fundamental results about modular forms that we will use. The reader should also be familiar with basic algebraic number theory, linear algebra, complex analysis (at the level of [Ahl78]), and algorithms (e.g., know what an algorithm is and what big oh notation means). In some of the examples and applications we assume that the reader knows about elliptic curves at the level of [Sil92]. Chapter 1 is foundational for the rest of this book. It introduces congruence subgroups of SL2 (Z) and modular forms as functions on the complex upper half plane. We discuss q-expansions, which provide an important computational handle on modular forms. We also study an algorithm for computing with congruence subgroups. The chapter ends with a list of applications of modular forms throughout mathematics. In Chapter 2 we discuss level 1 modular forms in much more detail. In particular, we introduce Eisenstein series and the cusp form ∆ and describe their q-expansions and basic properties. Then we prove a structure theorem for level 1 modular forms and use it to deduce dimension formulas and give an algorithm for explicitly computing a basis. We next introduce Hecke operators on level 1 modular forms, prove several results about them, and deduce multiplicativity of the Ramanujan τ function as an application. We also discuss explicit computation of Hecke operators. In Section 2.6 we make some brief remarks on recent work on asymptotically fast computation of values of τ . Finally, we describe computation of constant terms of Eisenstein series using an analytic algorithm. We generalize many of the constructions in this chapter to higher level in subsequent chapters. In Chapter 3 we turn to modular forms of higher level but restrict for simplicity to weight 2 since much is clearer in this case. (We remove the weight restriction later in Chapter 8.) We describe a geometric way of viewing cuspidal modular forms as differentials on modular curves, which leads to modular symbols, which are an explicit way to present a certain homology group. This chapter closes with methods for explicitly computing cusp forms of weight 2 using modular symbols, which we generalize in Chapter 9. In Chapter 4 we introduce Dirichlet characters, which are important both in explicit construction of Eisenstein series (in Chapter 5) and in decomposing spaces of modular forms as direct sums of simpler spaces. The

Preface

xiii

main focus of this chapter is a detailed study of how to explicitly represent and compute with Dirichlet characters. Chapter 5 is about how to explicitly construct the Eisenstein subspace of modular forms. First we define generalized Bernoulli numbers attached to a Dirichlet character and an integer then explain a new analytic algorithm for computing them (which generalizes the algorithm in Chapter 2). Finally we give without proof an explicit description of a basis of Eisenstein series, explain how to compute it, and give some examples. Chapter 6 records a wide range of dimension formulas for spaces of modular forms, along with a few remarks about where they come from and how to compute them. Chapter 7 is about linear algebra over exact fields, mainly the rational numbers. This chapter can be read independently of the others and does not require any background in modular forms. Nonetheless, this chapter occupies a central position in this book, because the algorithms in this chapter are of crucial importance to any actual implementation of algorithms for computing with modular forms. Chapter 8 is the most important chapter in this book; it generalizes Chapter 3 to higher weight and general level. The modular symbols formulation described here is central to general algorithms for computing with modular forms. Chapter 9 applies the algorithms from Chapter 8 to the problem of computing with modular forms. First we discuss decomposing spaces of modular forms using Dirichlet characters, and then explain how to compute a basis of Hecke eigenforms for each subspace using several approaches. We also discuss congruences between modular forms and bounds needed to provably generate the Hecke algebra. Chapter 10 is about computing analytic invariants of modular forms. It discusses tricks for speeding convergence of certain infinite series and sketches how to compute every elliptic curve over Q with given conductor. Chapter 11 contains detailed solutions to most of the exercises in this book. (Many of these were written by students in a course taught at the University of Washington.) Appendix A deals with computational techniques for working with generalizations of modular forms to more general groups than SL2 (Z), such as SLn (Z) for n ≥ 3. Some of this material requires more prerequisites than the rest of the book. Nonetheless, seeing a natural generalization of the material in the rest of this book helps to clarify the key ideas. The topics in the appendix are directly related to the main themes of this book: modular

xiv

Preface

symbols, Manin symbols, cohomology of subgroups of SL2 (Z) with various coefficients, explicit computation of modular forms, etc.

Software. We use SAGE, Software for Algebra and Geometry Experimentation (see [Ste06]), to illustrate how to do many of the examples. SAGE is completely free and packages together a wide range of open source mathematics software for doing much more than just computing with modular forms. SAGE can be downloaded and run on your computer or can be used via a web browser over the Internet. The reader is encouraged to experiment with many of the objects in this book using SAGE. We do not describe the basics of using SAGE in this book; the reader should read the SAGE tutorial (and other documentation) available at the SAGE website [Ste06]. All examples in this book have been automatically tested and should work exactly as indicated in SAGE version at least 1.5. Acknowledgements. David Joyner and Gabor Wiese carefully read the book and provided a huge number of helpful comments. John Cremona and Kevin Buzzard both made many helpful remarks that were important in the development of the algorithms in this book. Much of the mathematics (and some of the writing) in Chapter 10 is joint work with Helena Verrill. Noam Elkies made remarks about Chapters 1 and 2. S´andor Kov´ acs provided interesting comments on Chapter 1. Allan Steel provided helpful feedback on Chapter 7. Jordi Quer made useful remarks about Chapter 4 and Chapter 6. The students in the courses that I taught on this material at Harvard, San Diego, and Washington provided substantial feedback: in particular, Abhinav Kumar made numerous observations about computing widths of cusps (see Section 1.4.1) and Thomas James Barnet-Lamb made helpful remarks about how to represent Dirichlet characters. James Merryfield made helpful remarks about complex analytic issues and about convergence in Stirling’s formula. Robert Bradshaw, Andrew Crites (who wrote Exercise 7.5), Michael Goff, Dustin Moody, and Koopa Koo wrote most of the solutions included in Chapter 11 and found numerous typos throughout the book. Dustin Moody also carefully read through the book and provided feedback. H. Stark suggested using Stirling’s formula in Section 2.7.1, and Mark Watkins and Lynn Walling made comments on Chapter 3. Justin Walker found typos in the first published version of the book.

Preface

xv

Parts of Chapter 1 follow Serre’s beautiful introduction to modular forms [Ser73, Ch. VII] closely, though we adjust the notation, definitions, and order of presentation to be consistent with the rest of this book. I would like to acknowledge the partial support of NSF Grant DMS 0555776. Gunnells was supported in part by NSF Grants DMS 02-45580 and DMS 04-01525. Notation and Conventions. We denote canonical isomorphisms by ∼ = and noncanonical isomorphisms by ≈. If V is a vector space and s denotes some sort of construction involving V , we let Vs denote the corresponding subspace and V s the quotient space. E.g., if ι is an involution of V , then V+ is Ker(ι − 1) and V + = V /Im(ι − 1). If A is a finite abelian group, then Ator denotes the torsion subgroup and A/tor denotes the quotient A/Ator . We denote right group actions using exponential notation. Everywhere in this book, N is a positive integer and k is an integer. If N is an integer, a divisor t of N is a positive integer such that N/t is an integer.

Chapter 1

Modular Forms

This chapter introduces modular forms and congruence subgroups, which are central objects in this book. We first introduce the upper half plane and the group SL2 (Z) then recall some definitions from complex analysis. Next we define modular forms of level 1 followed by modular forms of general level. In Section 1.4 we discuss congruence subgroups and explain a simple way to compute generators for them and determine element membership. Section 1.5 lists applications of modular forms. We assume familiarity with basic number theory, group theory, and complex analysis. For a deeper understanding of modular forms, the reader is urged to consult the standard books in the field, e.g., [Lan95, Ser73, DI95, Miy89, Shi94, Kob84]. See also [DS05], which is an excellent first introduction to the theoretical foundations of modular forms.

1.1. Basic Definitions The group ¶ ¾ ½µ a b : ad − bc = 1 and a, b, c, d ∈ R SL2 (R) = c d acts on the complex upper half plane h = {z ∈ C : Im(z) > 0} by linear fractional transformations, as follows. If γ = for any z ∈ h we let (1.1.1)

γ(z) =

¡a b¢ c d

∈ SL2 (R), then

az + b ∈ h. cz + d 1

2

1. Modular Forms

Since the determinant of γ is 1, we have µ ¶ d 1 γ (z) = . dz (cz + d)2 Definition ¢ (Modular Group). The modular group is the group of all ¡ a b1.1 matrices c d with a, b, c, d ∈ Z and ad − bc = 1.

For example, the matrices ¶ µ 0 −1 (1.1.2) S= 1 0

and

T =

¶ µ 1 1 0 1

are both elements of SL2 (Z); the matrix S induces the function z 7→ −1/z on h, and T induces the function z 7→ z + 1. Theorem 1.2. The group SL2 (Z) is generated by S and T . Proof. See e.g. [Ser73, §VII.1].

¤

In SAGE we compute the group SL2 (Z) and its generators as follows: sage: G = SL(2,ZZ); G Modular Group SL(2,Z) sage: S, T = G.gens() sage: S [ 0 -1] [ 1 0] sage: T [1 1] [0 1] Definition 1.3 (Holomorphic and Meromorphic). Let R be an open subset of C. A function f : R → C is holomorphic if f is complex differentiable at every point z ∈ R, i.e., for each z ∈ R the limit f (z + h) − f (z) h→0 h

f ′ (z) = lim

exists, where h may approach 0 along any path. A function f : R → C∪{∞} is meromorphic if it is holomorphic except (possibly) at a discrete set S of points in R, and at each α ∈ S there is a positive integer n such that (z − α)n f (z) is holomorphic at α. The function f (z) = ez is a holomorphic function on C; in contrast, 1/(z − i) is meromorphic on C but not holomorphic since it has a pole at i. The function e−1/z is not even meromorphic on C.

1.2. Modular Forms of Level 1

3

Modular forms are holomorphic functions on h that transform in a particular way under a certain subgroup of SL2 (Z). Before defining general modular forms, we define modular forms of level 1.

1.2. Modular Forms of Level 1 Definition 1.4 (Weakly Modular Function). A weakly modular function ¡ a b ¢ of weight k ∈ Z is a meromorphic function f on h such that for all γ = c d ∈ SL2 (Z) and all z ∈ h we have (1.2.1)

f (z) = (cz + d)−k f (γ(z)).

The constant functions are weakly modular of weight 0. There are no nonzero weakly modular functions of odd weight (see Exercise 1.4), and it is not obvious that there are any weakly modular functions of even weight k ≥ 2 (but there are, as we will see!). The product of two weakly modular functions of weights k1 and k2 is a weakly modular function of weight k1 +k2 (see Exercise 1.3). When k is even, (1.2.1) has a possibly more conceptual interpretation; namely (1.2.1) is the same as f (γ(z))(d(γ(z)))k/2 = f (z)(dz)k/2 . Thus (1.2.1) simply says that the weight k “differential form” f (z)(dz)k/2 is fixed under the action of every element of SL2 (Z). By Theorem 1.2, the group SL2 (Z) is generated by the matrices S and T of (1.1.2), so to show that a meromorphic function f on h is a weakly modular function, all we have to do is show that for all z ∈ h we have (1.2.2)

f (z + 1) = f (z)

and

f (−1/z) = z k f (z).

Suppose f is a weakly modular function of weight k. expansion P A Fourier 2πinz , for all of f , if it exists, is a representation of f as f (z) = ∞ a e n=m n z ∈ h. Let q = q(z) = e2πiz , which we view as a holomorphic function on C. Let D′ be the open unit disk with the origin removed, and note that q defines a map h → D′ . By (1.2.2) we have f (z + 1) = f (z), so there is a function F : D′ → C such that F (q(z)) = f (z). This function F is a complex-valued function on D′ , but it may or may not be well behaved at 0. Suppose that F is well behaved at 0, in the sense that for some m ∈ Z and all q in a neighborhood of 0 we have the equality (1.2.3)

F (q) =

∞ X

n=m

an q n .

4

1. Modular Forms

If this is the case, we say that f is meromorphic at ∞. If, moreover, m ≥ 0, we say that f is holomorphic at ∞. We also call (1.2.3) the q-expansion of f about ∞. Definition 1.5 (Modular Function). A modular function of weight k is a weakly modular function of weight k that is meromorphic at ∞. Definition 1.6 (Modular Form). A modular form of weight k (and level 1) is a modular function of weight k that is holomorphic on h and at ∞. If f is a modular form, then there are numbers an such that for all z ∈ h, (1.2.4)

f (z) =

∞ X

an q n .

n=0

Proposition 1.7. The above series converges for all z ∈ h. Proof. The function f (q) is holomorphic on D, so its Taylor series converges absolutely in D. ¤ Since e2πiz → 0 as z → i∞, we set f (∞) = a0 . Definition 1.8 (Cusp Form). A cusp form of weight k (and level 1) is a modular form of weight k such that f (∞) = 0, i.e., a0 = 0. Let C[[q]] be the ring of all formal power series in q. If k = 2, then 1 dq dq = 2πiqdz, so dz = 2πi q . If f (q) is a cusp form of weight 2, then 2πif (z)dz = f (q)

dq f (q) = dq ∈ C[[q]]dq. q q

Thus the differential 2πif (z)dz is holomorphic at ∞, since q is a local parameter at ∞.

1.3. Modular Forms of Any Level In this section we define spaces of modular forms of arbitrary level. Definition 1.9 (Congruence Subgroup). A congruence subgroup of SL2 (Z) is any subgroup of SL2 (Z) that contains Γ(N ) = Ker(SL2 (Z) → SL2 (Z/N Z)) for some positive integer N . The smallest such N is the level of Γ. The most important congruence subgroups in this book are ¶ ¾ ¶ µ ¶ µ ½µ 1 ∗ a b a b (mod N ) ≡ ∈ SL2 (Z) : Γ1 (N ) = 0 1 c d c d

5

1.3. Modular Forms of Any Level

and

¶ ¶ µ ¶ µ ½µ ∗ ∗ a b a b ≡ ∈ SL2 (Z) : Γ0 (N ) = 0 ∗ c d c d

¾ (mod N ) ,

where ∗ means any element. Both groups have level N (see Exercise 1.6).

Let k be an integer. Define the weight k right of GL2 (Q) on the ¡ action ¢ set of all functions f : h → C as follows. If γ = ac db ∈ GL2 (Q), let (1.3.1)

(f [γ]k )(z) = det(γ)k−1 (cz + d)−k f (γ(z)).

Proposition 1.10. Formula (1.3.1) defines a right action of GL2 (Z) on the set of all functions f : h → C; in particular, f [γ1 γ2 ]k = (f [γ1 ]k )[γ2 ]k .

Proof. See Exercise 1.7.

¤

Definition 1.11 (Weakly Modular Function). A weakly modular function of weight k for a congruence subgroup Γ is a meromorphic function f : h → C such that f [γ]k = f for all γ ∈ Γ. A central object in the theory of modular forms is the set of cusps An element γ =

¡a b¢ c d

P1 (Q) = Q ∪ {∞}.

∈ SL2 (Z) acts on P1 (Q) by ( az+b if z 6= ∞, γ(z) = acz+d if z = ∞. c

Also, note that if the denominator c or cz + d is 0 above, then γ(z) = ∞ ∈ P1 (Q). The set of cusps for a congruence subgroup Γ is the set C(Γ) of Γ-orbits of P1 (Q). (We will often identify elements of C(Γ) with a representative element from the orbit.) For example, the lemma below asserts that if Γ = SL2 (Z), then there is exactly one orbit, so C(SL2 (Z)) = {[∞]}. Lemma 1.12. For any cusps α, β ∈ P1 (Q) there exists γ ∈ SL2 (Z) such that γ(α) = β. Proof. This is Exercise 1.8.

¤

Proposition 1.13. For any congruence subgroup Γ, the set C(Γ) of cusps is finite. Proof. This is Exercise 1.9.

¤

6

1. Modular Forms

See [DS05, §3.8] and Algorithm 8.12 below for more discussion of cusps and results relevant to their enumeration. In order to define modular forms for general congruence subgroups, we next explain what it means for a function to be holomorphic on the extended upper half plane h∗ = h ∪ P1 (Q). See [Shi94, §1.3–1.5] for a detailed description of the correct topology to consider on h∗ . In particular, a basis of neighborhoods for α ∈ Q is given by the sets {α} ∪ D, where D is an open disc in h that is tangent to the real line at α. Recall from Section 1.2 that a weakly modular P function f on SL2 (Z) is n holomorphic at ∞ if its q-expansion is of the form ∞ n=0 an q .

In order to make sense of holomorphicity of a weakly modular function f for an arbitrary congruence subgroup Γ at any α ∈ Q, we first prove a lemma.

Lemma 1.14. If f : h → C is a weakly modular function of weight k for a congruence subgroup Γ and if δ ∈ SL2 (Z), then f [δ]k is a weakly modular function for δ −1 Γδ. Proof. If s = δ −1 γδ ∈ δ −1 Γδ, then (f [δ]k )[s]k = f [δs]k = f [δδ

−1 γδ] k

= f [γδ]k = f [δ]k . ¤

Fix a weakly modular function f of weight k for a congruence subgroup Γ, and suppose α ∈ Q. In Section 1.2 we constructed the of ¶ µ q-expansion 1 1 ∈ SL2 (Z). f by using that f (z) = f (z + 1), which held since T = 0 1 There are congruence subgroups Γ such that T 6∈ Γ. Moreover, even if we are interested only in modular forms for Γ1 (N ), where we have T ∈ Γ1 (N ) for all N , we will still have to consider q-expansions at infinity for modular forms on δ −1 Γ1 (N )δ, and these need not contain T . Fortunately, ¡ groups ¢ N 1 N T = 0 1 ∈ Γ(N ), so a congruence subgroup of level N contains T N . Thus we have f (z + H) = f (H) for some positive integer H, e.g., H = N always works, but¡there ¢ may be a smaller choice of H. The minimal choice of H > 0 such that 10 H1 ∈ δ −1 Γδ, where δ(∞) = α, is called the width of the cusp α relative to the group Γ (see Section 1.4.1). When f is meromorphic at infinity, we obtain a Fourier expansion (1.3.2)

f (z) =

∞ X

n=m

an q n/H

1.4. Remarks on Congruence Subgroups

7

in powers of the function q 1/H = e2πiz/H . We say that f is holomorphic at ∞ if in (1.3.2) we have m ≥ 0. What about the other cusps α ∈ P1 (Q)? By Lemma 1.12 there is a γ ∈ SL2 (Z) such that γ(∞) = α. We declare f to be holomorphic at the cusp α if the weakly modular function f [γ]k is holomorphic at ∞.

Definition 1.15 (Modular Form). A modular form of integer weight k for a congruence subgroup Γ is a weakly modular function f : h → C that is holomorphic on h∗ . We let Mk (Γ) denote the space of weight k modular forms of weight k for Γ. Proposition 1.16. If a weakly modular function f is holomorphic at a set of representative elements for C(Γ), then it is holomorphic at every element of P1 (Q). Proof. Let c1 , . . . , cn ∈ P1 (Q) be representatives for the set of cusps for Γ. If α ∈ P1 (Q), then there is γ ∈ Γ such that α = γ(ci ) for some i. By hypothesis f is holomorphic at ci , so if δ ∈ SL2 (Z) is such that δ(∞) = ci , then f [δ]k is holomorphic at ∞. Since f is a weakly modular function for Γ, (1.3.3)

f [δ]k = (f [γ]k )[δ]k = f [γδ]k .

But γ(δ(∞)) = γ(ci ) = α, so (1.3.3) implies that f is holomorphic at α.

¤

1.4. Remarks on Congruence Subgroups Recall that a congruence subgroup is a subgroup of SL2 (Z) that contains Γ(N ) for some N . Any congruence subgroup has finite index in SL2 (Z), since Γ(N ) does. What about the converse: is every finite index subgroup of SL2 (Z) a congruence subgroup? This is the congruence subgroup problem. One can ask about the congruence subgroup problem with SL2 (Z) replaced by many similar groups. If p is a prime, then one can prove that every finite index subgroup of SL2 (Z[1/p]) is a congruence subgroup (i.e., contains the kernel of reduction modulo some integer coprime to p), and for any n > 2, all finite index subgroups of SLn (Z) are congruence subgroups (see [Hum80]). However, there are numerous finite index subgroups of SL2 (Z) that are not congruence subgroups. The paper [Hsu96] contains an algorithm to decide if certain finite index subgroups are congruence subgroups and gives an example of a subgroup of index 12 that is not a congruence subgroup. One can consider modular forms even for noncongruence subgroups. See, e.g., [Tho89] and the papers it references for work on this topic. We will not consider such modular forms further in this book. Note that modular symbols (which we define later in this book) are computable for noncongruence subgroups.

8

1. Modular Forms

Finding coset representatives for Γ0 (N ), Γ1 (N ) and Γ(N ) in SL2 (Z) is straightforward and will be discussed at length later in this book. To make the problem more explicit, note that you can quotient out by Γ(N ) first. Then the question amounts to finding coset representatives for a subgroup of SL2 (Z/N Z) (and lifting), which is reasonably straightforward. Given coset representatives for a finite index subgroup G of SL2 (Z), we can compute generators for G as follows. Let R be a set of coset representatives for G. Let σ, τ ∈ SL2 (Z) be the matrices denoted by S and T in (1.1.2). Define maps s, t : R → G as follows. If r ∈ R, then there exists a unique αr ∈ R such that Grσ = Gαr . Let s(r) = rσαr−1 . Likewise, there is a unique βr such that Grτ = Gβr and we let t(r) = rτ βr−1 . Note that s(r) and t(r) are in G for all r. Then G is generated by s(R) ∪ t(R). Proposition 1.17. The above procedure computes generators for G. Proof. Without loss of generality, assume that I = ( 10 01 ) represents the coset of G. Let g be an element of G. Since σ and τ generate SL2 (Z), it is possible to write g as a product of powers of σ and τ . There is a procedure, which we explain below with an example in order to avoid cumbersome notation, which writes g as a product of elements of s(R) ∪ t(R) times a right coset representative r ∈ R. For example, if g = στ 2 στ,

then g = Iστ 2 στ = s(I)yτ 2 στ for some y ∈ R. Continuing, s(I)yτ 2 στ = s(I)(yτ )τ στ = s(I)(t(y)z)τ στ

for some z ∈ R. Again, s(I)(t(y)z)τ στ = s(I)t(y)(zτ )στ = · · · . The procedure illustrated above (with an example) makes sense for arbitrary g and, after carrying it out, writes g as a product of elements of s(R) ∪ t(R) times a right coset representative r ∈ R. But g ∈ G and I is the right coset representative for G, so this right coset representative must be I. ¤ Remark 1.18. We could also apply the proof of Proposition 1.17 to write any element of G in terms of the given generators. Moreover, we could use it to write any element γ ∈ SL2 (Z) in the form gr, where g ∈ G and r ∈ R, so we can decide whether or not γ ∈ G. 1.4.1. Computing Widths of Cusps. Let Γ be a congruence subgroup of level N . Suppose α ∈ C(Γ) is a cusp, and choose (Z) such that ¡ γ ¢∈ SL2−1 1 h γ(∞) = α. Recall that the minimal h such that 0 1 ∈ γ Γγ is called the width of the cusp α for the group Γ. In this section we discuss how to compute h.

1.5. Applications of Modular Forms

9

Algorithm 1.19 (Width of Cusp). Given a congruence subgroup Γ of level N and a cusp α for Γ, this algorithm computes the width h of α. We assume that Γ is given by congruence conditions, e.g., Γ = Γ0 (N ) or Γ1 (N ). (1) [Find γ] Use the extended Euclidean algorithm to find γ ∈ SL2 (Z) such that γ(∞) = α, as follows. If α = ∞, set γ = 1; otherwise, ¡ ¢ write α = a/b, find c, d such that ad − bc = 1, and set γ = ac db . (2) [Compute Conjugate Matrix] Compute the following element of Mat2 (Z[x]): ¶ µ 1 x −1 γ . δ(x) = γ 0 1 Note that the entries of δ(x) are constant or linear in x. (3) [Solve] The congruence conditions that define Γ give rise to four linear congruence conditions on x. Use techniques from elementary number theory (or enumeration) to find the smallest simultaneous positive solution h to these four equations. Example 1.20. ¡ 0 −1 ¢ (1) Suppose α = 0 and Γ = Γ0 (N ) or Γ1 (N ). Then γ = 1 0 has the property that γ(∞) = α. Next, the congruence condition is ¶ ¶ µ ¶ µ µ 1 ∗ 1 0 1 x −1 (mod N ). ≡ γ = δ(x) = γ 0 1 −x 1 0 1

Thus the smallest positive solution is h = N , so the width of 0 is N .

(2) Suppose N¡ = ¢pq where p, q are distinct primes, and let α = 1/p. Then γ = p1 01 sends ∞ to α. The congruence condition for Γ0 (pq) is ¶ ¶ µ ¶ µ µ ∗ ∗ 1 − px x 1 x −1 (mod pq). ≡ γ = δ(x) = γ 0 ∗ −p2 x px + 1 0 1 Since p2 x ≡ 0 (mod pq), we see that x = q is the smallest solution. Thus 1/p has width q, and symmetrically 1/q has width p.

Remark 1.21. For Γ0 (N ), once we enforce that the bottom left entry is 0 (mod N ) and use that the determinant is 1, the coprimality from the other two congruences is automatic. So there is one congruence to solve in the Γ0 (N ) case. There are two congruences in the Γ1 (N ) case.

1.5. Applications of Modular Forms The above definition of modular forms might leave the impression that modular forms occupy an obscure corner of complex analysis. This is not the case! Modular forms are highly geometric, arithmetic, and topological objects that are of extreme interest all over mathematics:

10

1. Modular Forms

(1) Fermat’s last theorem: Wiles’ proof [Wil95] of Fermat’s last theorem uses modular forms extensively. The work of Wiles et al. on modularity also massively extends computational methods for elliptic curves over Q, because many elliptic curve algorithms, e.g., for computing L-functions, modular degrees, Heegner points, etc., require that the elliptic curve be modular. (2) Diophantine equations: Wiles’ proof of Fermat’s last theorem has made available a wide array of new techniques for solving certain diophantine equations. Such work relies crucially on having access to tables or software for computing modular forms. See, e.g., [Dar97, Mer99, Che05, SC03]. (Wiles did not need a computer, because the relevant spaces of modular forms that arise in his proof have dimension 0!) Also, according to Siksek (personal communication) the paper [BMS06] would “have been entirely impossible to write without [the algorithms described in this book].” (3) Congruent number problem: This ancient open problem is to determine which integers are the area of a right triangle with rational side lengths. There is a potential solution that uses modular forms (of weight 3/2) extensively (the solution is conditional on truth of the Birch and Swinnerton-Dyer conjecture, which is not yet known). See [Kob84]. (4) Topology: Topological modular forms are a major area of current research. (5) Construction of Ramanujan graphs: Modular forms can be used to construct almost optimal expander graphs, which play a role in communications network theory. (6) Cryptography and Coding Theory: Point counting on elliptic curves over finite fields is crucial to the construction of elliptic curve cryptosystems, and modular forms are relevant to efficient algorithms for point counting (see [Elk98]). Algebraic curves that are associated to modular forms are useful in constructing and studying certain error-correcting codes (see [Ebe02]). (7) The Birch and Swinnerton-Dyer conjecture: This central open problem in arithmetic geometry relates arithmetic properties of elliptic curves (and abelian varieties) to special values of L-functions. Most deep results toward this conjecture use modular forms extensively (e.g., work of Kolyvagin, Gross-Zagier, and Kato). Also, modular forms are used to compute and prove results about special values of these L-functions. See [Wil00].

11

1.6. Exercises

(8) Serre’s Conjecture on modularity of Galois representation: Let GQ = Gal(Q/Q) be the Galois group of an algebraic closure of Q. Serre conjectured and many people have (nearly!) proved that every continuous homomorphism ρ : GQ → GL2 (Fq ), where Fq is a finite field and det(ρ(complex conjugation)) = −1, “arises” from a modular form. More precisely, forPalmost all primes p the coefficients ap of a modular (eigen-)form an q n are congruent to the traces of elements ρ(Frobp ), where Frobp are certain special elements of GQ called Frobenius elements. See [RS01] and [DS05, Ch. 9]. (9) Generating functions for partitions: The generating functions for various kinds of partitions of an integer can often be related to modular forms. Deep theorems about modular forms then translate into results about partitions. See work of Ramanujan, Gordon, Andrews, and Ahlgren and Ono (e.g., [AO01]). (10) Lattices: If L ⊂ Rn is an even unimodular lattice (the basis matrix has determinant ±1 and λ · λ ∈ 2Z for all λ ∈ L), then the theta series X q λ·λ θL (q) = λ∈L

is a modular form of weight n/2. The coefficient of q m is the number of lattice vectors with squared length m. Theorems and computational methods for modular forms translate into theorems and computational methods for lattices. For example, the 290 theorem of M. Bharghava and J. Hanke is a theorem about lattices, which asserts that an integer-valued quadratic form represents all positive integers if and only if it represents the integers up to 290; it is proved by doing many calculations with modular forms (both theoretical and with a computer).

1.6. Exercises ¡ ¢ 1.1 Suppose γ = ac db ∈ GL2 (R) has positive determinant. Prove that if z ∈ C is a complex number with positive imaginary part, then the imaginary part of γ(z) = (az + b)/(cz + d) is also positive. 1.2 Prove that every rational function (quotient of two polynomials) is a meromorphic function on C. 1.3 Suppose f and g are weakly modular functions for a congruence subgroup Γ with f 6= 0. (a) Prove that the product f g is a weakly modular function for Γ. (b) Prove that 1/f is a weakly modular function for Γ.

12

1. Modular Forms

(c) If f and g are modular functions, show that f g is a modular function for Γ. (d) If f and g are modular forms, show that f g is a modular form for Γ. 1.4 Suppose f is a weakly modular function of odd weight k and level Γ0 (N ) for some N . Show that f = 0. 1.5 Prove that SL2 (Z) = Γ0 (1) = Γ1 (1) = Γ(1). 1.6 (a) Prove that Γ1 (N ) is a group. (b) Prove that Γ1 (N ) has finite index in SL2 (Z) (Hint: It contains the kernel of the homomorphism SL2 (Z) → SL2 (Z/N Z).) (c) Prove that Γ0 (N ) has finite index in SL2 (Z). (d) Prove that Γ0 (N ) and Γ1 (N ) have level N . 1.7 ¡Let ¢k be an integer, and for any function f : h∗ → C and γ = a b ∈ GL2 (Q), set f [γ]k (z) = det(γ)k−1 · (cz + d)−k · f (γ(z)). c d Prove that if γ1 , γ2 ∈ GL2 (Z), then for all z ∈ h∗ we have f [γ1 γ2 ]k (z) = ((f [γ1 ]k )[γ2 ]k )(z).

1.8 Prove that for any α, β ∈ P1 (Q), there exists γ ∈ SL2 (Z) such that γ(α) = β. 1.9 Prove Proposition 1.13, which asserts that the set of cusps C(Γ), for any congruence subgroup Γ, is finite. 1.10 Use Algorithm 1.19 to give an example of a group Γ and cusp α with width 2.

Chapter 2

Modular Forms of Level 1

In this chapter we study in detail the structure of level 1 modular forms, i.e., modular forms on SL2 (Z) = Γ0 (1) = Γ1 (1). We assume some complex analysis (e.g., the residue theorem), linear algebra, and that the reader has read Chapter 1.

2.1. Examples of Modular Forms of Level 1 In this section we will finally see some examples of modular forms of level 1! We first introduce the Eisenstein series and then define ∆, which is a cusp form of weight 12. In Section 2.2 we prove the structure theorem, which says that all modular forms of level 1 are polynomials in Eisenstein series. For an even integer k ≥ 4, the nonnormalized weight k Eisenstein series is the function on the extended upper half plane h∗ = h ∪ P1 (Q) given by (2.1.1)

Gk (z) =

∗ X

m,n∈Z

1 . (mz + n)k

The star on top of the sum symbol means that for each z the sum is over all m, n ∈ Z such that mz + n 6= 0. Proposition 2.1. The function Gk (z) is a modular form of weight k, i.e., Gk ∈ Mk (SL2 (Z)). 13

14

2. Modular Forms of Level 1

Proof. See [Ser73, § VII.2.3] for a proof that Gk (z) defines a holomorphic function on h∗ . To see that Gk is modular, observe that Gk (z + 1) =

∗ X

∗

∗

X X 1 1 1 = = , k k (m(z + 1) + n) (mz + (n + m)) (mz + n)k

where for the last equality we use that the map (m, n + m) 7→ (m, n) on Z × Z is invertible. Also, Gk (−1/z) = =

∗ X

1 (−m/z + n)k

∗ X

= zk

zk (−m + nz)k ∗ X 1

(mz + n)k

= z k Gk (z),

where we use that (n, −m) 7→ (m, n) is invertible.

¤

Proposition 2.2. Gk (∞) = 2ζ(k), where ζ is the Riemann zeta function. Proof. As z → ∞ (along the imaginary axis) in (2.1.1), the terms that involve z with m 6= 0 go to 0. Thus Gk (∞) = This sum is twice ζ(k) =

∗ X 1 . nk

n∈Z

P

1 n≥1 nk ,

as claimed.

¤

2.1.1. The Cusp Form ∆. Suppose E = C/Λ is an elliptic curve over C, viewed as a quotient of C by a lattice Λ = Zω1 + Zω2 , with ω1 /ω2 ∈ h (see [DS05, §1.4]). The Weierstrass ℘-function of the lattice Λ is X 1 (k − 1)Gk (ω1 /ω2 )uk−2 , ℘ = ℘Λ (u) = 2 + u k=4,6,8,...

where the sum is over even integers k ≥ 4. It satisfies the differential equation (℘′ )2 = 4℘3 − 60G4 (ω1 /ω2 )℘ − 140G6 (ω1 /ω2 ).

If we set x = ℘ and y = ℘′ , the above is an (affine) equation of the form y 2 = ax3 +bx+c for an elliptic curve that is complex analytically isomorphic to C/Λ (see [Ahl78, pg. 277] for why the cubic has distinct roots). The discriminant of the cubic 4x3 − 60G4 (ω1 /ω2 )x − 140G6 (ω1 /ω2 ) is 16D(ω1 /ω2 ), where D(z) = (60G4 (z))3 − 27(140G6 (z))2 .

2.1. Examples of Modular Forms of Level 1

15

Since D(z) is the difference of two modular forms of weight 12 it has weight 12. Moreover, D(∞) = (60G4 (∞))3 − 27 (140G6 (∞))2 µ ¶ ¶ µ 60 4 3 140 · 2 6 2 = π π − 27 32 · 5 33 · 5 · 7 = 0, so D is a cusp form of weight 12. Let ∆=

D . (2π)12

Lemma 2.3. If z ∈ h, then ∆(z) 6= 0. Proof. Let ω1 = z and ω2 = 1. Since E = C/(Zω1 + Zω2 ) is an elliptic curve, it has nonzero discriminant ∆(z) = ∆(ω1 /ω2 ) 6= 0. ¤ Q∞ Proposition 2.4. We have ∆ = q · n=1 (1 − q n )24 . Proof. See [Ser73, Thm. 6, pg. 95].

¤

Remark 2.5. SAGE computes the q-expansion of ∆ efficiently to high precision using the command delta qexp: sage: delta_qexp(6) q - 24*q^2 + 252*q^3 - 1472*q^4 + 4830*q^5 + O(q^6) 2.1.2. Fourier Expansions of Eisenstein Series. Recall from (1.2.4) that elements f of Mk (SL2 (Z)) can be expressed as formal power series in terms of q(z) = e2πiz and that this expansion is called the Fourier expansion of f . The following proposition gives the Fourier expansion of the Eisenstein series Gk (z). Definition 2.6 (Sigma). For any integer t ≥ 0 and any positive integer n, the sigma function X dt σt (n) = 1≤d|n

is the sum of the tth powers of the positive divisors of n. Also, let d(n) = σ0 (n), which is the number of divisors of n, and let σ(n) = σ1 (n). For example, if p is prime, then σt (p) = 1 + pt . Proposition 2.7. For every even integer k ≥ 4, we have ∞ (2πi)k X σk−1 (n)q n . · Gk (z) = 2ζ(k) + 2 · (k − 1)! n=1

16

2. Modular Forms of Level 1

Proof. See [Ser73, Section VII.4], which uses clever manipulations of series, starting with the identity ¶ ∞ µ 1 X 1 1 π cot(πz) = + . + z z+m z−m m=1

¤ From a computational point of view, the q-expansion of Proposition 2.7 is unsatisfactory because it involves transcendental numbers. To understand these numbers, we introduce the Bernoulli numbers Bn for n ≥ 0 defined by the following equality of formal power series: ∞

X xn x B = . n ex − 1 n!

(2.1.2)

n=0

Expanding the power series, we have x x x2 x4 x6 x8 = 1 − + − + − + ··· . ex − 1 2 12 720 30240 1209600

As this expansion suggests, the Bernoulli numbers Bn with n > 1 odd are 0 (see Exercise 1.2). Expanding the series further, we obtain the following table: 1 1 1 1 1 B0 = 1, B1 = − , B2 = , B4 = − , B6 = , B8 = − , 2 6 30 42 30 B10 =

5 , 66

B20 = −

B12 = −

174611 , 330

691 , 2730

B22 =

7 B14 = , 6

854513 , 138

B16 = −

B24 = −

3617 , 510

236364091 , 2730

B18 =

43867 , 798

B26 =

8553103 . 6

See Section 2.7 for a discussion of fast (analytic) methods for computing Bernoulli numbers. We compute some Bernoulli numbers in SAGE: sage: bernoulli(12) -691/2730 sage: bernoulli(50) 495057205241079648212477525/66 sage: len(str(bernoulli(10000))) 27706 A key fact is that Bernoulli numbers are rational numbers and they are connected to values of ζ at positive even integers.

2.2. Structure Theorem for Level 1 Modular Forms

17

Proposition 2.8. If k ≥ 2 is an even integer, then ζ(k) = −

(2πi)k · Bk . 2 · k!

Proof. This is proved by manipulating a series expansion of z cot(z) (see [Ser73, Section VII.4]). ¤ Definition 2.9 (Normalized Eisenstein Series). The normalized Eisenstein series of even weight k ≥ 4 is Ek =

(k − 1)! · Gk . 2 · (2πi)k

Combining Propositions 2.7 and 2.8, we see that Ek = −

(2.1.3)

∞

X Bk σk−1 (n)q n . +q+ 2k n=2

Warning 2.10. Our series Ek is normalized so that the coefficient of q is 1, but often in the literature Ek is normalized so that the constant coefficient is 1. We use the normalization with the coefficient of q equal to 1, because then the eigenvalue of the nth Hecke operator (see Section 2.4) is the coefficient of q n . Our normalization is also convenient when considering congruences between cusp forms and Eisenstein series.

2.2. Structure Theorem for Level 1 Modular Forms In this section we describe a structure theorem for modular forms of level 1. If f is a nonzero meromorphic function on h and w ∈ h, let ordw (f ) be the largest integer n such that f (z)/(w − z)n is holomorphic at w. If f = P ∞ n n=m an q with am 6= 0, we set ord∞ (f ) = m. We will use the following theorem to give a presentation for the vector space of modular forms of weight k; this presentation yields an algorithm to compute this space. Let Mk = Mk (SL2 (Z)) denote the complex vector space of modular forms of weight k for SL2 (Z). The standard fundamental domain F for SL2 (Z) is the set of z ∈ h with |z| ≥ 1 and |Re(z)| ≤ 1/2. Let ρ = e2πi/3 . Theorem 2.11 (Valence Formula). Let k be any integer and suppose f ∈ Mk (SL2 (Z)) is nonzero. Then ∗ X 1 k 1 ordw (f ) = , ord∞ (f ) + ordi (f ) + ordρ (f ) + 2 3 12 w∈F

where

∗ X

w∈F

is the sum over elements of F other than i and ρ.

18

2. Modular Forms of Level 1

Proof. The proof in [Ser73, §VII.3] uses the residue theorem.

¤

Let Sk = Sk (SL2 (Z)) denote the subspace of weight k cusp forms for SL2 (Z). We have an exact sequence ι

∞ 0 → Sk → Mk −− →C

that sends f ∈ Mk to f (∞). When k ≥ 4 is even, the space Mk contains the Eisenstein series Gk , and Gk (∞) = 2ζ(k) 6= 0, so the map Mk → C is surjective. This proves the following lemma. Lemma 2.12. If k ≥ 4 is even, then Mk = Sk ⊕ CGk and the following sequence is exact: ι∞ 0 → Sk → Mk −− → C → 0. Proposition 2.13. For k < 0 and k = 2, we have Mk = 0. Proof. Suppose f ∈ Mk is nonzero yet k = 2 or k < 0. By Theorem 2.11, ∗ X k 1 1 1 ordw (f ) = ord∞ (f ) + ordi (f ) + ordρ (f ) + ≤ . 2 3 12 6 w∈D

This is not possible because each quantity on the left is nonnegative so whatever the sum is, it is too big (or 0, in which case k = 0). ¤

Theorem 2.14. Multiplication by ∆ defines an isomorphism Mk−12 → Sk . Proof. By Lemma 2.3, ∆ is not identically 0, so because ∆ is holomorphic, multiplication by ∆ defines an injective map Mk−12 ֒→ Sk . To see that this map is surjective, we show that if f ∈ Sk , then f /∆ ∈ Mk−12 . Since ∆ has weight 12 and ord∞ (∆) ≥ 1, Theorem 2.11 implies that ∆ has a simple zero at ∞ and does not vanish on h. Thus if f ∈ Sk and if we let g = f /∆, then g is holomorphic and satisfies the appropriate transformation formula, so g ∈ Mk−12 . ¤ Corollary 2.15. For k = 0, 4, 6, 8, 10, 14, the space Mk has dimension 1, with basis 1, G4 , G6 , G8 , G10 , and G14 , respectively, and Sk = 0. Proof. Combining Proposition 2.13 with Theorem 2.14, we see that the spaces Mk for k ≤ 10 cannot have dimension greater than 1, since otherwise Mk′ 6= 0 for some k ′ < 0. Also M14 has dimension at most 1, since M2 has dimension 0. Each of the indicated spaces of weight ≥ 4 contains the indicated Eisenstein series and so has dimension 1, as claimed. ¤   if k is odd or negative, 0 Corollary 2.16. dim Mk = ⌊k/12⌋ if k ≡ 2 (mod 12),   ⌊k/12⌋ + 1 if k 6≡ 2 (mod 12). Here ⌊x⌋ is the biggest integer ≤ x.

2.2. Structure Theorem for Level 1 Modular Forms

19

Proof. As we have already seen above, the formula is true when k ≤ 12. By Theorem 2.14, the dimension increases by 1 when k is replaced by k+12. ¤ Theorem 2.17. The space Mk has as basis the modular forms Ga4 Gb6 , where a, b run over all pairs of nonnegative integers such that 4a + 6b = k. Proof. Fix an even integer k. We first prove by induction that the modular forms Ga4 Gb6 generate Mk ; the cases k ≤ 10 and k = 14 follow from the above arguments (e.g., when k = 0, we have a = b = 0 and basis 1). Choose some pair of nonnegative integers a, b such that 4a + 6b = k. The form g = Ga4 Gb6 is not a cusp form, since it is nonzero at ∞. Now suppose f ∈ Mk is arbitrary. Since g(∞) 6= 0, there exists α ∈ C such that f − αg ∈ Sk . Then by Theorem 2.14, there is h ∈ Mk−12 such that f − αg = ∆ · h. By induction, h is a polynomial in G4 and G6 of the required type, and so is ∆, so f is as well. Thus {Ga4 Gb6 | a ≥ 0, b ≥ 0, 4a + 6b = k} spans Mk . Suppose there is a nontrivial linear relation between the Ga4 Gb6 for a given k. By multiplying the linear relation by a suitable power of G4 and G6 , we may assume that we have such a nontrivial relation with k ≡ 0 (mod 12). k/6 Now divide the linear relation by the weight k form G6 to see that G34 /G26 satisfies a polynomial with coefficients in C (see Exercise 2.4). Hence G34 /G26 is a root of a polynomial, hence a constant, which is a contradiction since the q-expansion of G34 /G26 is not constant. ¤ Algorithm 2.18 (Basis for Mk ). Given integers n and k, this algorithm computes a basis of q-expansions for the complex vector space Mk mod q n . The q-expansions output by this algorithm have coefficients in Q. (1) [Simple Case] If k = 0, output the basis with just 1 in it and terminate; otherwise if k < 4 or k is odd, output the empty basis and terminate. (2) [Power Series] Compute E4 and E6 mod q n using the formula from (2.1.3) and Section 2.7. (3) [Initialize] Set b = 0. (4) [Enumerate Basis] For each integer b between 0 and ⌊k/6⌋, compute a = (k − 6b)/4. If a is an integer, compute and output the basis element E4a E6b mod q n . When computing E4a , find E4m (mod q n ) for each m ≤ a, and save these intermediate powers, so they can be reused later, and likewise for powers of E6 .

20

2. Modular Forms of Level 1

Proof. This is simply a translation of Theorem 2.17 into an algorithm, since Ek is a nonzero scalar multiple of Gk . That the q-expansions have coefficients in Q follows from (2.1.3). ¤ Example 2.19. We compute a basis for M24 , which is the space with smallest weight whose dimension is greater than 1. It has as basis E46 , E43 E62 , and E64 , whose explicit expansions are 1 1 203 + q+ q2 + · · · , 191102976000000 132710400000 44236800000 1 1 377 E43 E62 = − q− q2 + · · · , 3511517184000 12192768000 4064256000 1 241 1 − q+ q2 + · · · . E64 = 64524128256 32006016 10668672 E46 =

We compute this basis in SAGE as follows: sage: E4 = eisenstein_series_qexp(4, 3) sage: E6 = eisenstein_series_qexp(6, 3) sage: E4^6 1/191102976000000 + 1/132710400000*q + 203/44236800000*q^2 + O(q^3) sage: E4^3*E6^2 1/3511517184000 - 1/12192768000*q - 377/4064256000*q^2 + O(q^3) sage: E6^4 1/64524128256 - 1/32006016*q + 241/10668672*q^2 + O(q^3) In Section 2.3, we will discuss the reduced echelon form basis for Mk .

2.3. The Miller Basis Lemma 2.20 (V. Miller). The space Sk has a basis f1 , . . . , fd such that if ai (fj ) is the ith coefficient of fj , then ai (fj ) = δi,j for i = 1, . . . , d. Moreover the fj all lie in Z[[q]]. We call this basis the Miller basis for Sk . This is a straightforward construction involving E4 , E6 and ∆. The following proof very closely follows [Lan95, Ch. X, Thm. 4.4], which in turn follows the first lemma of V. Miller’s thesis. Proof. Let d = dim Sk . Since B4 = −1/30 and B6 = 1/42, we note that F4 = −

8 · E4 = 1 + 240q + 2160q 2 + 6720q 3 + 17520q 4 + · · · B4

21

2.3. The Miller Basis

and F6 = −

12 · E6 = 1 − 504q − 16632q 2 − 122976q 3 − 532728q 4 + · · · B6

have q-expansions in Z[[q]] with leading coefficient 1. Choose integers a, b ≥ 0 such that 4a + 6b ≤ 14

4a + 6b ≡ k

and

(mod 12),

with a = b = 0 when k ≡ 0 (mod 12), and let gj = ∆

j

2(d−j)+b a F6 F4

=

µ

∆ F62

¶j

F62d+b F4a ,

for j = 1, . . . , d.

Then it is elementary to check that gj has weight k aj (gj ) = 1

and

ai (gj ) = 0

when

i < j.

Hence the gj are linearly independent over C, so form a basis for Sk . Since F4 , F6 , and ∆ are all in Z[[q]], so are the gj . The fi may then be constructed from the gj by Gauss elimination. The coefficients of the resulting power series lie in Z because each time we clear a column we use the power series gj whose leading coefficient is 1 (so no denominators are introduced). ¤ Remark 2.21. The basis coming from Miller’s lemma is “canonical”, since it is just the reduced row echelon form of any basis. Also the set of all integral linear combinations of the elements of the Miller basis are precisely the modular forms of level 1 with integral q-expansion. We extend the Miller basis to all Mk by taking a multiple of Gk with constant term 1 and subtracting off the fi from the Miller basis so that the coefficients of q, q 2 , . . . q d of the resulting expansion are 0. We call the extra basis element f0 . Example 2.22. If k = 24, then d = 2. Choose a = b = 0, since k ≡ 0 (mod 12). Then g1 = ∆F62 = q − 1032q 2 + 245196q 3 + 10965568q 4 + 60177390q 5 − · · · and g2 = ∆2 = q 2 − 48q 3 + 1080q 4 − 15040q 5 + · · · . We let f2 = g2 and f1 = g1 + 1032g2 = q + 195660q 3 + 12080128q 4 + 44656110q 5 − · · · .

22

2. Modular Forms of Level 1

Example 2.23. When k = 36, the Miller basis including f0 is f0 = 1 + f1 = f2 = f3 =

q+

6218175600q 4 + 15281788354560q 5 + · · · , 57093088q 4 +

q2 +

194184q 4 +

q3 −

72q 4 +

37927345230q 5 + · · · , 7442432q 5 + · · · ,

2484q 5 + · · · .

Example 2.24. The SAGE command victor miller basis computes the Miller basis to any desired precision for a given k. sage: victor_miller_basis(28,5) [ 1 + 15590400*q^3 + 36957286800*q^4 + O(q^5), q + 151740*q^3 + 61032448*q^4 + O(q^5), q^2 + 192*q^3 - 8280*q^4 + O(q^5) ] Remark 2.25. To write f ∈ Mk as a polynomial in E4 and E6 , it is wasteful to compute the Miller basis. Instead, use the upper triangular (but not 2(d−j)+a b echelon!) basis ∆j F6 F4 , and match coefficients from q 0 to q d .

2.4. Hecke Operators In this section we define Hecke operators on level 1 modular forms and derive their basic properties. We will not give proofs of the analogous properties for Hecke operators on higher level modular forms, since the proofs are clearest in the level 1 case, and the general case is similar (see, e.g., [Lan95]). For any positive integer n, let ¶ ¾ ½µ a b ∈ Mat2 (Z) : a ≥ 1, ad = n, and 0 ≤ b < d . Xn = 0 d

2 Note that the ¡set X ¢ n is in bijection with the set of subgroups of Z of a b index n, where c d corresponds to L = Z · (a, b) + Z · (0, d), as one can see using Hermite normal form, which is the analogue over Z of echelon form (see Exercise 7.5). ¡ ¢ Recall from (1.3.1) that if γ = ac db ∈ GL2 (Q), then

f [γ]k = det(γ)k−1 (cz + d)−k f (γ(z)).

Definition 2.26 (Hecke Operator Tn,k ). The nth Hecke operator Tn,k of weight k is the operator on the set of functions on h defined by X Tn,k (f ) = f [γ]k . γ∈Xn

23

2.4. Hecke Operators

Remark 2.27. It would make more sense to write Tn,k on the right, e.g., f |Tn,k , since Tn,k is defined using a right group action. However, if n, m are integers, then the action of Tn,k and Tm,k on weakly modular functions commutes (by Proposition 2.29 below), so it makes no difference whether we view the Hecke operators of given weight k as acting on the right or left. Proposition 2.28. If f is a weakly modular function of weight k, then so is Tn,k (f ); if f is a modular function, then so is Tn,k (f ). Proof. Suppose γ ∈ SL2 (Z). Since γ induces an automorphism of Z2 , Xn · γ = {δγ : δ ∈ Xn }

is also in bijection with the subgroups of Z2 of index n. For each element δγ ∈ Xn · γ, there is σ ∈ SL2 (Z) such that σδγ ∈ Xn (the element σ transforms δγ to Hermite normal form), and the set of elements σδγ is thus equal to Xn . Thus X X f [δγ]k = Tn,k (f )[γ]k . f [σδγ]k = Tn,k (f ) = σδγ∈Xn

δ∈Xn

A finite sum of meromorphic function is meromorphic, so Tn,k (f ) is weakly modular. If f is holomorphic on h, then each f [δ]k is holomorphic on h for δ ∈ Xn . A finite sum of holomorphic functions is holomorphic, so Tn,k (f ) is holomorphic. ¤ We will frequently drop k from the notation in Tn,k , since the weight k is implicit in the modular function to which we apply the Hecke operator. Henceforth we make the convention that if we write Tn (f ) and if f is modular, then we mean Tn,k (f ), where k is the weight of f . Proposition 2.29. On weight k modular functions we have (2.4.1)

Tmn = Tm Tn

if (m, n) = 1,

and (2.4.2)

Tpn = Tpn−1 Tp − pk−1 Tpn−2

if p is prime.

Proof. Let L be a subgroup of index mn. The quotient Z2 /L is an abelian group of order mn, and (m, n) = 1, so Z2 /L decomposes uniquely as a direct sum of a subgroup of order m with a subgroup of order n. Thus there exists a unique subgroup L′ such that L ⊂ L′ ⊂ Z2 , and L′ has index m in Z2 . The subgroup L′ corresponds to an element of Xm , and the index n subgroup L ⊂ L′ corresponds to multiplying that element on the right by some uniquely determined element of Xn . We thus have SL2 (Z) · Xm · Xn = SL2 (Z) · Xmn ,

24

2. Modular Forms of Level 1

i.e., the set products of elements in Xm with elements of Xn equal the elements of Xmn , up to SL2 (Z)-equivalence. Thus for any f , we have Tmn (f ) = Tn (Tm (f )). Applying this formula with m and n swapped yields the equality Tmn = Tm Tn . We will show that Tpn + pk−1 Tpn−2 = Tp Tpn−1 . Suppose f is a weight k “ ” p0 [ ]

weakly modular function. Using that f 0 p k = (p2 )k−1 p−k f = pk−2 f , we have X X X X f [x]k . f [x]k = f [x]k + p f [x]k + pk−1 x∈Xpn

Also Tp Tpn−1 (f ) =

x∈Xpn−2

X

X

x∈Xpn

(f [x]k )[y]k =

y∈Xp x∈Xpn−1

x∈pXpn−2

X

f [x]k .

x∈Xpn−1 ·Xp

Thus it suffices to show that Xpn disjoint union p copies of pXpn−2 is equal to Xpn−1 · Xp , where we consider elements with multiplicities and up to left SL2 (Z)-equivalence (i.e., the left action of SL2 (Z)). Suppose L is a subgroup of Z2 of index pn , so L corresponds to an element of Xpn . First suppose L is not contained in pZ2 . Then the image of L in Z2 /pZ2 = (Z/pZ)2 is of order p, so if L′ = pZ2 + L, then [Z2 : L′ ] = p and [L : L′ ] = pn−1 , and L′ is the only subgroup with this property. Second, suppose that L ⊂ pZ2 if of index pn and that x ∈ Xpn corresponds to L. Then every one of the p + 1 subgroups L′ ⊂ Z2 of index p contains L. Thus there are p + 1 chains L ⊂ L′ ⊂ Z2 with [Z2 : L′ ] = p.

The chains L ⊂ L′ ⊂ Z2 with [Z2 : L′ ] = p and [Z2 : L] = pn−1 are in bijection with the elements of Xpn−1 · Xp . On the other hand the union of Xpn with p copies of pXpn−2 corresponds to the subgroups L of index pn , but with those that contain pZ2 counted p + 1 times. The structure of the set of chains L ⊂ L′ ⊂ Z2 that we derived in the previous paragraph gives the result. ¤ Corollary 2.30. The Hecke operator Tpn , for prime p, is a polynomial in Tp with integer coefficients, i.e., Tpn ∈ Z[Tp ]. If n, m are any integers, then Tn T m = Tm Tn . Proof. The first statement follows from (2.4.2) of Proposition 2.29. It then follows that Tn Tm = Tm Tn when m and n are both powers of a single prime p. Combining this with (2.4.1) gives the second statement in general. ¤ P Proposition 2.31. Let f = n∈Z an q n be a modular function of weight k. Then   X X  dk−1 amn/d2  q m . Tn (f ) = m∈Z

1≤d | gcd(n,m)

25

2.4. Hecke Operators

In particular, if n = p is prime, then ´ X³ Tp (f ) = amp + pk−1 am/p q m , m∈Z

where am/p = 0 if m/p 6∈ Z. Proof. This is proved in [Ser73, §VII.5.3] by writing out Tn (f ) explicitly P ¤ and using that 0≤b 1 and ε > 0, find M ∈ Z+ so that Y z= (1 − p−s )−1 p≤M

satisfies 0 ≤ ζ(s) − z < ε. We always have 0 ≤ ζ(s) − z. Also, X Y n−s ≤ (1 − p−s )−1 , n≤M

so

ζ(s) − z ≤

∞ X

n=M +1

p≤M

n

−s

≤

Z

∞

x−s dx =

M

1 . (s − 1)M s−1

Thus if M > ε−1/(s−1) , then 1 1 ≤ s−1 < ε , s−1 (s − 1)M M

so ζ(s) − z < ε, as required. For our purposes, we have s = n and ε = (Kd)−1 , so it suffices to take M > (Kd)1/(n−1) . Algorithm 2.45 (Bernoulli Number Bn ). Given an integer n ≥ 0, this algorithm computes the Bernoulli number Bn as an exact rational number. (1) [Special cases] If n = 0, return 1; if n = 1, return −1/2; if n ≥ 3 is odd, return 0. 2 · n! (2) [Factorial factor] Compute K = to sufficiently many digits (2π)n of precision so the ceiling in step (6) is uniquely determined (this precision can be determined using Y Section 2.7.1). p. (3) [Denominator] Compute d = p−1|n l m (4) [Bound] Compute M = (Kd)1/(n−1) . Y (5) [Approximate ζ(n)] Compute z = (1 − p−n )−1 . p≤M n/2+1

(6) [Numerator] Compute a = (−1) a (7) [Output Bn ] Return . d

⌈dKz⌉.

In step (5) use a sieve to compute all primes p ≤ M efficiently (which is fast, since M is so small). In step (4) we may replace M by any integer greater than the one specified by the formula, so we do not have to compute (Kd)1/(n−1) to very high precision. In Section 5.2.2 below we will generalize the above algorithm.

33

2.8. Exercises

Example 2.46. We illustrate Algorithm 2.45 by computing B50 . Using 135 binary digits of precision, we compute K = 7500866746076957704747736.71552473164563479. The divisors of n are 1, 2, 5, 10, 25, 50, so d = 2 · 3 · 11 = 66. We find M = 4 and compute z = 1.00000000000000088817842109308159029835012. Finally we compute dKz = 495057205241079648212477524.999999994425778, so B50 =

495057205241079648212477525 . 66

2.8. Exercises 2.1 Using Proposition 2.8 and the table on page 16, compute explicitly.

P∞

1 n=1 n26

2.2 Prove that if n > 1 is odd, then the Bernoulli number Bn is 0. 2.3 Use (2.1.3) to write down the coefficients of 1, q, q 2 , and q 3 of the Eisenstein series E8 . 2.4 Suppose k is a positive integer with k ≡ 0 (mod 12). Suppose a, b ≥ 0 are integers with 4a + 6b = k. (a) Prove 3 | a. k ¢a ¡ (b) Show that Ga4 · Gb6 / G66 = G34 /G26 3 .

2.5 Compute the Miller basis for M28 (SL2 (Z)) with precision O(q 8 ). Your answer will look like Example 2.23. 2.6 Consider the cusp form f = q 2 +192q 3 −8280q 4 +· · · in S28 (SL2 (Z)). Write f as a polynomial in E4 and E6 (see Remark 2.25).

2.7 Let Gk be the weight k Eisenstein series from equation (2.1.1). Let c be the complex number so that the constant coefficient of the q-expansion of g = c · Gk is 1. Is it always the case that the q-expansion of g lies in Z[[q]]? 2.8 Compute the matrix of the Hecke operator T2 on the Miller basis for M32 (SL2 (Z)). Then compute its characteristic polynomial and verify it factors as a product of two irreducible polynomials. What Next? Much of the rest of this book is about methods for computing subspaces of Mk (Γ1 (N )) for general N and k. These general methods are

34

2. Modular Forms of Level 1

more complicated than the methods presented in this chapter, since there are many more modular forms of small weight and it can be difficult to obtain them. Forms of level N > 1 have subtle connections with elliptic curves, abelian varieties, and motives. Read on for more!

Chapter 3

Modular Forms of Weight 2

We saw in Chapter 2 (especially Section 2.2) that we can compute each space Mk (SL2 (Z)) explicitly. This involves computing Eisenstein series E4 and E6 to some precision, then forming the basis {E4a E6b : 4a + 6b = k, 0 ≤ a, b ∈ Z} for Mk (SL2 (Z)). In this chapter we consider the more general problem of computing S2 (Γ0 (N )), for any positive integer N . Again we have a decomposition M2 (Γ0 (N )) = S2 (Γ0 (N )) ⊕ E2 (Γ0 (N )), where E2 (Γ0 (N )) is spanned by generalized Eisenstein series and S2 (Γ0 (N )) is the space of cusp forms, i.e., elements of M2 (Γ0 (N )) that vanish at all cusps. In Chapter 5 we compute the space E2 (Γ0 (N )) in a similar way to how we computed Mk (SL2 (Z)). On the other hand, elements of S2 (Γ0 (N )) often cannot be written as sums or products of generalized Eisenstein series. In fact, the structure of M2 (Γ0 (N )) is, in general, much more complicated than that of Mk (SL2 (Z)). For example, when p is a prime, E2 (Γ0 (p)) has dimension 1, whereas S2 (Γ0 (p)) has dimension about p/12. Fortunately an idea of Birch, which he called modular symbols, provides a method for computing S2 (Γ0 (N )) and indeed for much more that is relevant to understanding special values of L-functions. Modular symbols are also a powerful theoretical tool. In this chapter, we explain how S2 (Γ0 (N )) is related to modular symbols and how to use this relationship to explicitly 35

36

3. Modular Forms of Weight 2

compute a basis for S2 (Γ0 (N )). In Chapter 8 we will introduce more general modular symbols and explain how to use them to compute Sk (Γ0 (N )), Sk (Γ1 (N )) and Sk (N, ε) for any integers k ≥ 2 and N and character ε.

Section 3.1 contains a very brief summary of basic facts about modular forms of weight 2, modular curves, Hecke operators, and integral homology. Section 3.2 introduces modular symbols and describes how to compute with them. In Section 3.5 we talk about how to cut out the subspace of modular symbols corresponding to cusp forms using the boundary map. Section 3.6 is about a straightforward method to compute a basis for S2 (Γ0 (N )) using modular symbols, and Section 3.7 outlines a more sophisticated algorithm for computing newforms that uses Atkin-Lehner theory. Before reading this chapter, you should have read Chapter 1 and Chapter 2. We also assume familiarity with algebraic curves, Riemann surfaces, and homology groups of compact Riemann surfaces.

3.1. Hecke Operators Recall from Chapter 1 that the group Γ0 (N ) acts on h∗ = h∪P1 (Q) by linear fractional transformations. The quotient Γ0 (N )\h∗ is a Riemann surface, which we denote by X0 (N ). See [DS05, Ch. 2] for a detailed description of the topology on X0 (N ). The Riemann surface X0 (N ) also has a canonical structure of algebraic curve over Q, as is explained in [DS05, Ch. 7] (see also [Shi94, §6.7]). Recall from Section 1.3 that a cusp form of weight 2 for Γ0 (N ) is a function f on h such that f (z)dz defines a holomorphic differential on X0 (N ). Equivalently, a cusp form is a holomorphic function f on h such that

(a) the expression f (z)dz is invariant under replacing z by γ(z) for each γ ∈ Γ0 (N ) and

(b) f (z) vanishes at every cusp for Γ0 (N ).

The space S2 (Γ0 (N )) of weight 2 cusp forms on Γ0 (N ) is a finite-dimensional complex vector space, of dimension equal to the genus g of X0 (N ). The space X0 (N )(C) is a compact oriented Riemann surface, so it is a 2-dimensional oriented real manifold, i.e., X0 (N )(C) is a g-holed torus (see Figure 3.1.1 on page 38). Condition (b) in the definition of f means that f has a Fourier expansion about each element of P1 (Q). Thus, at ∞ we have f (z) = a1 e2πiz + a2 e2πi2z + a3 e2πi3z + · · · = a1 q + a2 q 2 + a3 q 3 + · · · ,

where, for brevity, we write q = q(z) = e2πiz .

37

3.1. Hecke Operators

Example 3.1. Let E be the elliptic curve defined by the equation y 2 +xy = ˜ p ), where E ˜ is the reduction of E x3 + x2 − 4x − 5. Let ap = p + 1 − #E(F mod p (note that for the primes that divide the conductor of E we have a3 = −1, a13 = 1). For n composite, define an using the relations at the end of Section 3.7. Then the Shimura-Taniyama conjecture asserts that f = q + a2 q 2 + a3 q 3 + a4 q 4 + a5 q 5 + · · · = q + q 2 − q 3 − q 4 + 2q 5 + · · ·

is the q-expansion of an element of S2 (Γ0 (39)). This conjecture, which is now a theorem (see [BCDT01]), asserts that any q-expansion constructed as above from an elliptic curve over Q is a modular form. This conjecture was mostly proved first by Wiles [Wil95] as a key step in the proof of Fermat’s last theorem. Just as is the case for level 1 modular forms (see Section 2.4) there are commuting Hecke operators T1 , T2 , T3 , . . . that act on S2 (Γ0 (N )). To define them conceptually, we introduce an interpretation of the modular curve X0 (N ) as an object whose points parameterize elliptic curves with extra structure. Proposition 3.2. The complex points of Y0 (N ) = Γ0 (N )\h are in natural bijection with isomorphism classes of pairs (E, C), where E is an elliptic curve over C and C is a cyclic subgroup of E(C) of order N . The class of the point λ ∈ h corresponds to the pair ¶ ¶ µ µ 1 Z + Zλ /(Z + Zλ) . C/(Z + Zλ), N Proof. See Exercise 3.1.

¤

Suppose n and N are coprime positive integers. There are two natural maps π1 and π2 from Y0 (n · N ) to Y0 (N ); the first, π1 , sends (E, C) ∈ Y0 (n · N )(C) to (E, C ′ ), where C ′ is the unique cyclic subgroup of C of order N , and the second, π2 , sends (E, C) to (E/D, C/D), where D is the unique cyclic subgroup of C of order n. These maps extend in a unique way to algebraic maps from X0 (n · N ) to X0 (N ): (3.1.1)

X0 (n · N )

r rrr r r r xx rr r π2

X0 (N )

LLL LLLπ1 LLL L&&

X0 (N ).

The nth Hecke operator Tn is π1∗ ◦π2∗ , where π2∗ and π1∗ denote pullback and pushforward of differentials, respectively. (There is a similar definition of Tn when gcd(n, N ) 6= 1.) Using our interpretation of S2 (Γ0 (N )) as differentials

38

3. Modular Forms of Weight 2

on X0 (N ), this gives an action of Hecke operators on S2 (Γ0 (N )). One can show that these induce the maps of Proposition 2.31 on q-expansions. Example 3.3. There is a basis of S2 (39) so that     1 1 0 −4 −2 −6 4 4 . T2 = −2 −3 −2 and T5 =  4 0 0 1 0 0 2

Notice that these matrices commute. Also, the characteristic polynomial of T2 is (x − 1) · (x2 + 2x − 1). 3.1.1. Homology. The first homology group H1 (X0 (N ), Z) is the group of closed 1-cycles modulo boundaries of 2-cycles (formal sums of images of 2-simplexes). Topologically X0 (N ) is a g-holed torus, where g is the genus of X0 (N ). Thus H1 (X0 (N ), Z) is a free abelian group of rank 2g (see, e.g., [GH81, Ex. 19.30] and [DS05, §6.1]), with two generators corresponding to each hole, as illustrated in the case N = 39 in Figure 3.1.1.

H1 (X0 (39), Z) ∼ =Z×Z×Z×Z×Z×Z Figure 3.1.1. The homology of X0 (39).

The homology of X0 (N ) is closely related to modular forms, since the Hecke operators Tn also act on H1 (X0 (N ), Z). The action is by pullback of homology classes by π2 followed by taking the image under π1 , where π1 and π2 are as in (3.1.1). Integration defines a pairing (3.1.2)

h , i : S2 (Γ0 (N )) × H1 (X0 (N ), Z) → C.

Explicitly, for a path x,

hf, xi = 2πi ·

Z

f (z)dz.

x

Theorem 3.4. The pairing (3.1.2) is nondegenerate and Hecke equivariant in the sense that for every Hecke operator Tn , we have hf Tn , xi = hf, Tn xi. Moreover, it induces a perfect pairing (3.1.3)

h , i : S2 (Γ0 (N )) × H1 (X0 (N ), R) → C.

39

3.2. Modular Symbols

This is a special case of the results in Section 8.5. As we will see, modular symbols allow us to make explicit the action of the Hecke operators on H1 (X0 (N ), Z); the above pairing then translates this into a wealth of information about cusp forms. We will also consider the relative homology group H1 (X0 (N ), Z; {cusps}) of X0 (N ) relative to the cusps; it is the same as usual homology, but in addition we allow paths with endpoints in the cusps instead of restricting to closed loops. Modular symbols provide a “combinatorial” presentation of H1 (X0 (N ), Z) in terms of paths between elements of P1 (Q).

3.2. Modular Symbols Let M2 be the free abelian group with basis the set of symbols {α, β} with α, β ∈ P1 (Q) modulo the 3-term relations {α, β} + {β, γ} + {γ, α} = 0

above and modulo any torsion. Since M2 is torsion-free, we have {α, α} = 0

and

{α, β} = −{β, α}.

Remark 3.5 (Warning). The symbols {α, β} satisfy the relations {α, β} = −{β, α}, so order matters. The notation {α, β} looks like the set containing two elements, which strongly (and incorrectly) suggests that the order does not matter. This is the standard notation in the literature.

Figure 3.2.1. The modular symbols {α, β} and {0, ∞}.

As illustrated in Figure 3.2.1, we “think of” this modular symbol as the homology class, relative to the cusps, of a path from α to β in h∗ .

40

3. Modular Forms of Weight 2

Define a left action of GL2 (Q) on M2 by letting g ∈ GL2 (Q) act by g{α, β} = {g(α), g(β)}, and g acts on α and β via the corresponding linear fractional transformation. The space M2 (Γ0 (N )) of modular symbols for Γ0 (N ) is the quotient of M2 by the submodule generated by the infinitely many elements of the form x − g(x), for x in M2 and g in Γ0 (N ), and modulo any torsion. A modular symbol for Γ0 (N ) is an element of this space. We frequently denote the equivalence class of a modular symbol by giving a representative element. Example 3.6. Some modular symbols are 0 no matter what the level N is! For example, since γ = ( 10 11 ) ∈ Γ0 (N ), we have {∞, 0} = {γ(∞), γ(0)} = {∞, 1}, so 0 = {∞, 1} − {∞, 0} = {∞, 1} + {0, ∞} = {0, ∞} + {∞, 1} = {0, 1}. See Exercise 3.2 for a generalization of this observation. There is a natural homomorphism (3.2.1)

ϕ : M2 (Γ0 (N )) → H1 (X0 (N ), {cusps}, Z)

that sends a formal linear combination of geodesic paths in the upper half plane to their image as paths on X0 (N ). In [Man72] Manin proved that (3.2.1) is an isomorphism (this is a fairly involved topological argument). Manin identified the subspace of M2 (Γ0 (N )) that is sent isomorphically onto H1 (X0 (N ), Z). Let B2 (Γ0 (N )) denote the free abelian group whose basis is the finite set C(Γ0 (N )) = Γ0 (N )\P1 (Q) of cusps for Γ0 (N ). The boundary map δ : M2 (Γ0 (N )) → B2 (Γ0 (N ))

sends {α, β} to {β}−{α}, where {β} denotes the basis element of B2 (Γ0 (N )) corresponding to β ∈ P1 (Q). The kernel S2 (Γ0 (N )) of δ is the subspace of cuspidal modular symbols. Thus an element of S2 (Γ0 (N )) can be thought of as a linear combination of paths in h∗ whose endpoints are cusps and whose images in X0 (N ) are homologous to a Z-linear combination of closed paths. Theorem 3.7 (Manin). The map ϕ above induces a canonical isomorphism S2 (Γ0 (N )) ∼ = H1 (X0 (N ), Z). Proof. This is [Man72, Thm. 1.9]. For any (commutative) ring R let M2 (Γ0 (N ), R) = M2 (Γ0 (N )) ⊗Z R

¤

3.3. Computing with Modular Symbols

41

and S2 (Γ0 (N ), R) = S2 (Γ0 (N )) ⊗Z R. Proposition 3.8. We have dimC S2 (Γ0 (N ), C) = 2 dimC S2 (Γ0 (N )). Proof. We have dimC S2 (Γ0 (N ), C) = rankZ S2 (Γ0 (N )) = rankZ H1 (X0 (N ), Z) = 2g. ¤ Example 3.9. We illustrate modular symbols in the case when N = 11. Using SAGE (below), which implements the algorithm that we describe below over Q, we find that M2 (Γ0 (11); Q) has basis {∞, 0}, {−1/8, 0}, {−1/9, 0}. A basis for the integral homology H1 (X0 (11), Z) is the subgroup generated by {−1/8, 0} and {−1/9, 0}. sage: set_modsym_print_mode (’modular’) sage: M = ModularSymbols(11, 2) sage: M.basis() ({Infinity,0}, {-1/8,0}, {-1/9,0}) sage: S = M.cuspidal_submodule() sage: S.integral_basis() # basis over ZZ. ({-1/8,0}, {-1/9,0}) sage: set_modsym_print_mode (’manin’) # set it back

3.3. Computing with Modular Symbols 3.3.1. Manin’s Trick. In this section, we describe a trick of Manin that we will use to prove that spaces of modular symbols are computable. By Exercise 1.6 the group Γ0 (N ) has finite index in SL2 (Z). Fix right coset representatives r0 , r1 , . . . , rm for Γ0 (N ) in SL2 (Z), so that SL2 (Z) = Γ0 (N )r0 ∪ Γ0 (N )r1 ∪ · · · ∪ Γ0 (N )rm , where the union is disjoint. For example, when N is prime, a list of coset representatives is ¶ ¶ µ ¶ µ ¶ µ ¶ µ ¶ µ µ 0 −1 1 0 1 0 1 0 1 0 1 0 . , ,..., , , , 1 0 N −1 1 3 1 2 1 1 1 0 1 Let (3.3.1)

P1 (Z/N Z) = {(a : b) : a, b ∈ Z/N Z, gcd(a, b, N ) = 1 }/ ∼

where (a : b) ∼ (a′ : b′ ) if there is u ∈ (Z/N Z)∗ such that a = ua′ , b = ub′ .

42

3. Modular Forms of Weight 2

Proposition 3.10. There is a bijection between P1 (Z/N Z) and ¢ right ¡ the cosets of Γ0 (N ) in SL2 (Z), which sends a coset representative ac db to the class of (c : d) in P1 (Z/N Z). Proof. See Exercise 3.3.

¤

See Proposition 8.6 for the analogous statement for Γ1 (N ). We now describe an observation of Manin (see [Man72, §1.5]) that is crucial to making M2 (Γ0 (N )) computable. It allows us to write any modular symbol {α, β} as a Z-linear combination of symbols of the form ri {0, ∞}, where the ri ∈ SL2 (Z) are coset representatives as above. In particular, the finitely many symbols r0 {0, ∞}, . . . , rm {0, ∞} generate M2 (Γ0 (N )). Proposition 3.11 (Manin). Let N be a positive integer and r0 , . . . , rm a set of right coset representatives for Γ0 (N ) in SL2 (Z). Every {α, β} ∈ M2 (Γ0 (N )) is a Z-linear combination of r0 {0, ∞}, . . . , rm {0, ∞}. We give two proofs of the proposition. The first is useful for computation (see [Cre97a, §2.1.6]); the second (see [MTT86, §2]) is easier to understand conceptually since it does not require any knowledge of continued fractions. Continued Fractions Proof of Proposition 3.11. Since {α, β} = {0, β} − {0, α}, it suffices to consider modular symbols of the form {0, b/a}, where the rational number b/a is in lowest terms. Expand b/a as a continued fraction and consider the successive convergents in lowest terms: 0 b−2 = , a−2 1

b−1 1 = , a−1 0

b0 b0 = , ..., a0 1

bn−1 , an−1

bn b = an a

where the first two are included formally. Then bk ak−1 − bk−1 ak = (−1)k−1 , so that gk = Hence

µ

bk (−1)k−1 bk−1 ak (−1)k−1 ak−1

¶

∈ SL2 (Z).

½

¾ bk−1 bk , = gk {0, ∞} = ri {0, ∞}, ak−1 ak for some i, is of the required special form. Since ¾ ½ ¾ ½ bn−1 bn b0 b1 , , + ··· + , {0, b/a} = {0, ∞} + {∞, b0 } + 1 a1 an−1 an this completes the proof.

¤

43

3.3. Computing with Modular Symbols

Inductive Proof of Proposition 3.11. As in the first proof it suffices to prove the proposition for any symbol {0, b/a}, where b/a is in lowest terms. We will induct on a ∈ Z≥0 . If a = 0, then the symbol is {0, ∞}, which corresponds to the identity coset, so assume that a > 0. Find a′ ∈ Z such that ba′ ≡ 1 (mod a); then b′ = (ba′ − 1)/a ∈ Z so the matrix µ ¶ b b′ δ= a a′

is an element of SL2 (Z). Thus δ = γ · rj for some right coset representative rj and γ ∈ Γ0 (N ). Then ¶ µ b b′ ′ ′ ′ ′ · {0, ∞} = rj {0, ∞}, {0, b/a} − {0, b /a } = {b /a , b/a} = a a′

as elements of M2 (Γ0 (N )). By induction, {0, b′ /a′ } is a linear combination of symbols of the form rk {0, ∞}, which completes the proof. ¤ Example 3.12. Let N = 11, and consider the modular symbol {0, 4/7}. We have 1 4 , =0+ 7 1 + 11 1+ 3

so the partial convergents are b−2 0 = , a−2 1

b−1 1 = , a−1 0

b0 0 = , a0 1

b1 1 = , a1 1

b2 1 = , a2 2

b3 4 = . a3 7

Thus, noting as in Example 3.6 that {0, 1} = 0, we have {0, 4/7} = {0, ∞} + {∞, 0} + {0, 1} + {1, 1/2} + {1/2, 4/7} ¶ ¶ µ µ 4 1 1 −1 {0, ∞} {0, ∞} + = 7 2 2 −1 µ ¶ µ ¶ 1 0 1 0 = {0, ∞} + {0, ∞} 9 1 9 1 ·µ ¶ ¸ 1 0 = 2· {0, ∞} . 9 1 We compute the convergents of 4/7 in SAGE as follows (note that 0 and ∞ are excluded): sage: convergents(4/7) [0, 1, 1/2, 4/7]

44

3. Modular Forms of Weight 2

3.3.2. Manin Symbols. As above, fix coset representatives r0 , . . . , rm for Γ0 (N ) in SL2 (Z). Consider formal symbols [ri ]′ for i = 0, . . . , m. Let [ri ] be the modular symbol ri {0, ∞} = {ri (0), ri (∞)}. We equip the symbols [r0 ]′ , . . . , [rm ]′ with a right action of SL2 (Z), which is given by [ri ]′ .g = [rj ]′ , where Γ0 (N )rj = Γ0 (N )ri g. We extend the notation by writing [γ]′ = [Γ0 (N )γ]′ = [ri ]′ , where γ ∈ Γ0 (N )ri . Then the right action of Γ0 (N ) is simply [γ]′ .g = [γg]′ . Theorem 1.2¡ implies ¢ that SL2 (Z) is generated by the two matrices σ = ¢ ¡ 0 −1 1 −1 1 0 and τ = 1 0 . Note that σ = S from Theorem 1.2 and τ = T S, so T = τ σ ∈ hσ, τ i.

The following theorem provides us with a finite presentation for the space M2 (Γ0 (N )) of modular symbols. Theorem 3.13 (Manin). Consider the quotient M of the free abelian group on Manin symbols [r0 ]′ , . . . , [rm ]′ by the subgroup generated by the elements (for all i): [ri ]′ + [ri ]′ σ

and

[ri ]′ + [ri ]′ τ + [ri ]′ τ 2 ,

and modulo any torsion. Then there is an isomorphism ∼

Ψ : M −→ M2 (Γ0 (N )) given by [ri ]′ 7→ [ri ] = ri {0, ∞}. Proof. We will only prove that Ψ is surjective; the proof that Ψ is injective requires much more work and will be omitted from this book (see [Man72, §1.7] for a complete proof).

Proposition 3.11 implies that Ψ is surjective, assuming that Ψ is well defined. We next verify that Ψ is well defined, i.e., that the listed 2-term and 3-term relations hold in the image. To see that the first relation holds, note that [ri ] + [ri ]σ = {ri (0), ri (∞)} + {ri σ(0), ri σ(∞)} = {ri (0), ri (∞)} + {ri (∞), ri (0)}

= 0.

For the second relation we have [ri ] + [ri ]τ + [ri ]τ 2 = {ri (0), ri (∞)} + {ri τ (0), ri τ (∞)} + {ri τ 2 (0), ri τ 2 (∞)} = {ri (0), ri (∞)} + {ri (∞), ri (1)} + {ri (1), ri (0)}

= 0.

¤

3.3. Computing with Modular Symbols

45

Example 3.14. By default SAGE computes modular symbols spaces over Q, i.e., M2 (Γ0 (N ); Q) ∼ = M2 (Γ0 (N )) ⊗ Q. SAGE represents (weight 2) Manin symbols as pairs (c, d). Here c, d are integers that satisfy 0 ≤ c, d < N ; they define a point (c : d) ∈ P1 (Z/N Z), hence a right coset of Γ0 (N ) in SL2 (Z) (see Proposition 3.10). Create M2 (Γ0 (N ); Q) in SAGE by typing ModularSymbols(N, 2). We then use the SAGE command manin generators to enumerate a list of generators [r0 ], . . . , [rn ] as in Theorem 3.13 for several spaces of modular symbols.

sage: M = ModularSymbols(2,2) sage: M Modular Symbols space of dimension 1 for Gamma_0(2) of weight 2 with sign 0 over Rational Field sage: M.manin_generators() [(0,1), (1,0), (1,1)] sage: M = ModularSymbols(3,2) sage: M.manin_generators() [(0,1), (1,0), (1,1), (1,2)] sage: M = ModularSymbols(6,2) sage: M.manin_generators() [(0,1), (1,0), (1,1), (1,2), (1,3), (1,4), (1,5), (2,1), (2,3), (2,5), (3,1), (3,2)] Given x=(c,d), the command x.lift to sl2z(N) computes an element of SL2 (Z) whose lower two entries are congruent to (c, d) modulo N . sage: M = ModularSymbols(2,2) sage: [x.lift_to_sl2z(2) for x in M.manin_generators()] [[1, 0, 0, 1], [0, -1, 1, 0], [0, -1, 1, 1]] sage: M = ModularSymbols(6,2) sage: x = M.manin_generators()[9] sage: x (2,5) sage: x.lift_to_sl2z(6) [1, 2, 2, 5] The manin basis command returns a list of indices into the Manin generator list such that the corresponding symbols form a basis for the quotient

46

3. Modular Forms of Weight 2

of the Q-vector space spanned by Manin symbols modulo the 2-term and 3-term relations of Theorem 3.13. sage: M = ModularSymbols(2,2) sage: M.manin_basis() [1] sage: [M.manin_generators()[i] for i in M.manin_basis()] [(1,0)] sage: M = ModularSymbols(6,2) sage: M.manin_basis() [1, 10, 11] sage: [M.manin_generators()[i] for i in M.manin_basis()] [(1,0), (3,1), (3,2)] Thus, e.g., every element of M2 (Γ0 (6)) is a Q-linear combination of the three symbols [(1, 0)], [(3, 1)], and [(3, 2)]. We can write each of these as a modular symbol using the modular symbol rep function. sage: M.basis() ((1,0), (3,1), (3,2)) sage: [x.modular_symbol_rep() for x in M.basis()] [{Infinity,0}, {0,1/3}, {-1/2,-1/3}] The manin gens to basis function returns a matrix whose rows express each Manin symbol generator in terms of the subset of Manin symbols that forms a basis (as returned by manin basis). sage: M = ModularSymbols(2,2) sage: M.manin_gens_to_basis() [-1] [ 1] [ 0] Since the basis is (1, 0), this means that in M2 (Γ0 (2); Q), we have [(0, 1)] = −[(1, 0)] and [(1, 1)] = 0. (Since no denominators are involved, we have in fact computed a presentation of M2 (Γ0 (2); Z).) To convert a Manin symbol x = (c, d) to an element of a modular symbols space M , use M(x):

47

3.4. Hecke Operators

sage: M = ModularSymbols(2,2) sage: x = (1,0); M(x) (1,0) Next consider M2 (Γ0 (6); Q): sage: M = ModularSymbols(6,2) sage: M.manin_gens_to_basis() [-1 0 0] [ 1 0 0] [ 0 0 0] [ 0 -1 1] [ 0 -1 0] [ 0 -1 1] [ 0 0 0] [ 0 1 -1] [ 0 0 -1] [ 0 1 -1] [ 0 1 0] [ 0 0 1] Recall that our choice of basis for M2 (Γ0 (6); Q) is [(1, 0)], [(3, 1)], [(3, 2)]. Thus, e.g., the first row of this matrix says that [(0, 1)] = −[(1, 0)], and the fourth row asserts that [(1, 2)] = −[(3, 1)] + [(3, 2)]. sage: M = ModularSymbols(6,2) sage: M((0,1)) -(1,0) sage: M((1,2)) -(3,1) + (3,2)

3.4. Hecke Operators 3.4.1. Hecke Operators on Modular Symbols. When p is a prime not dividing N , define ¶ µ X µ1 r ¶ p 0 {α, β}. {α, β} + Tp ({α, β}) = 0 p 0 1 r mod p

The Hecke operators are compatible with the integration pairing h , i of Section 3.1, in the sense that hf Tp , xi = hf, Tp xi. When p | N , the definition

48

3. Modular Forms of Weight 2

¡ ¢ is the same, except that the matrix p0 10 is not included in the sum (see Theorem 8.23). There is a similar definition of Tn for n composite (see Section 8.3.1). Example 3.15. For example, when N = 11, we have T2 {0, 1/5} = {0, 2/5} + {0, 1/10} + {1/2, 3/5} = −2{0, 1/5}.

See Figure 3.4.1.

Figure 3.4.1. Image of {0, 1/5} under T2

3.4.2. Hecke Operators on Manin Symbols. In [Mer94], L. Merel gives a description of the action of Tp directly on Manin symbols [ri ] (see Section 8.3.2 for details). For example, when p = 2 and N is odd, we have ¶ ¶ µ ¶ µ ¶ µ µ 1 0 2 1 2 0 1 0 . + [ri ] + [ri ] + [ri ] (3.4.1) T2 ([ri ]) = [ri ] 1 2 0 1 0 1 0 2 For any prime, let Cp be the set of matrices constructed using the following algorithm (see [Cre97a, §2.4]): Algorithm 3.16 (Cremona’s Heilbronn Matrices). Given a prime p, this algorithm outputs a list of 2 × 2 matrices of determinant p that can be used to compute the Hecke operator Tp . ¶ µ 1 0 . (1) Output 0 p l pm jpk (2) For r = − , . . . , : 2 2 (a) Let x1 =µp, x2 =¶−r, y1 = 0, y2 = 1, a = −p, b = r. x1 x2 . (b) Output y1 y2 (c) As long as b 6= 0, do the following: (i) Let q be the integer closest to a/b (if a/b is a half integer, round away from 0). (ii) Let c = a − bq, a = −b, b = c.

49

3.4. Hecke Operators

(iii) Set x3 = qx2 − x1 , x1 = x2 , x2 = x3 , and y3 = qy2µ− y1 , y1¶= y2 , y2 = y3 . x1 x2 . (iv) Output y1 y2 Proposition 3.17 (Cremona, Merel). Let Cp be as above. Then for p ∤ N and [x] ∈ M2 (Γ0 (N )) a Manin symbol, we have X Tp ([x]) = [xg]. g∈Cp

Proof. See Proposition 2.4.1 of [Cre97a].

¤

There are other lists of matrices, due to Merel, that work even when p | N (see Section 8.3.2).

The command HeilbronnCremonaList(p), for p prime, outputs the list of matrices from Algorithm 3.16. sage: HeilbronnCremonaList(2) [[1, 0, 0, 2], [2, 0, 0, 1], [2, 1, 0, 1], sage: HeilbronnCremonaList(3) [[1, 0, 0, 3], [3, 1, 0, 1], [1, 0, 1, 3], [3, -1, 0, 1], [-1, 0, 1, -3]] sage: HeilbronnCremonaList(5) [[1, 0, 0, 5], [5, 2, 0, 1], [2, 1, 1, 3], [5, 1, 0, 1], [1, 0, 1, 5], [5, 0, 0, 1], [-1, 0, 1, -5], [5, -2, 0, 1], [-2, 1, 1, [1, 0, -3, 5]] sage: len(HeilbronnCremonaList(37)) 128 sage: len(HeilbronnCremonaList(389)) 1892 sage: len(HeilbronnCremonaList(2003)) 11662

[1, 0, 1, 2]] [3, 0, 0, 1],

[1, 0, 3, 5], [5, -1, 0, 1], -3],

Example 3.18. We compute the matrix of T2 on M2 (Γ0 (2)): sage: M = ModularSymbols(2,2) sage: M.T(2).matrix() [1]

Example 3.19. We compute some Hecke operators on M2 (Γ0 (6)):

50

3. Modular Forms of Weight 2

sage: M = ModularSymbols(6, 2) sage: M.T(2).matrix() [ 2 1 -1] [-1 0 1] [-1 -1 2] sage: M.T(3).matrix() [3 2 0] [0 1 0] [2 2 1] sage: M.T(3).fcp() # factored characteristic polynomial (x - 3) * (x - 1)^2 For p ≥ 5 we have Tp = p + 1, since M2 (Γ0 (6)) is spanned by generalized Eisenstein series (see Chapter 5). Example 3.20. We compute the Hecke operators on M2 (Γ0 (39)): sage: M = ModularSymbols(39, 2) sage: T2 = M.T(2) sage: T2.matrix() [ 3 0 -1 0 0 1 1 -1 0] [ 0 0 2 0 -1 1 0 1 -1] [ 0 1 0 -1 1 1 0 1 -1] [ 0 0 1 0 0 1 0 1 -1] [ 0 -1 2 0 0 1 0 1 -1] [ 0 0 1 1 0 1 1 -1 0] [ 0 0 0 -1 0 1 1 2 0] [ 0 0 0 1 0 0 2 0 1] [ 0 0 -1 0 0 0 1 0 2] sage: T2.fcp() # factored characteristic polynomial (x - 3)^3 * (x - 1)^2 * (x^2 + 2*x - 1)^2 The Hecke operators commute, so their eigenspace structures are related. sage: T2 = M.T(2).matrix() sage: T5 = M.T(5).matrix() sage: T2*T5 - T5*T2 == 0 True sage: T5.charpoly().factor() (x^2 - 8)^2 * (x - 6)^3 * (x - 2)^2

3.5. Computing the Boundary Map

51

The decomposition of T2 is a list of the kernels of (f e )(T2 ), where f runs through the irreducible factors of the characteristic polynomial of T2 and f e exactly divides this characteristic polynomial. Using SAGE, we find them: sage: M = ModularSymbols(39, 2) sage: M.T(2).decomposition() [ Modular Symbols subspace of dimension 3 of Modular Symbols space of dimension 9 for Gamma_0(39) of weight 2 with sign 0 over Rational Field, Modular Symbols subspace of dimension 2 of Modular Symbols space of dimension 9 for Gamma_0(39) of weight 2 with sign 0 over Rational Field, Modular Symbols subspace of dimension 4 of Modular Symbols space of dimension 9 for Gamma_0(39) of weight 2 with sign 0 over Rational Field ]

3.5. Computing the Boundary Map In Section 3.2 we defined a map δ : M2 (Γ0 (N )) → B2 (Γ0 (N )). The kernel of this map is the space S2 (Γ0 (N )) of cuspidal modular symbols. This kernel will be important in computing cusp forms in Section 3.7 below. To ¡compute the boundary map on [γ], note that [γ] = {γ(0), γ(∞)}, so ¢ a b if γ = c d , then δ([γ]) = {γ(∞)} − {γ(0)} = {a/c} − {b/d}.

Computing this boundary map would appear to first require an algorithm to compute the set C(Γ0 (N )) = Γ0 (N )\P1 (Q) of cusps for Γ0 (N ). In fact, there is a trick that computes the set of cusps in the course of running the algorithm. First, give an algorithm for deciding whether or not two elements of P1 (Q) are equivalent modulo the action of Γ0 (N ). Then simply construct C(Γ0 (N )) in the course of computing the boundary map, i.e., keep a list of cusps found so far, and whenever a new cusp class is discovered, add it to the list. The following proposition, which is proved in [Cre97a, Prop. 2.2.3], explains how to determine whether two cusps are equivalent. Proposition 3.21 (Cremona). Let (ci , di ), i = 1, 2, be pairs of integers with gcd(ci , di ) = 1 and possibly di = 0. There is g ∈ Γ0 (N ) such that g(c1 /d1 ) = c2 /d2 in P1 (Q) if and only if s1 d2 ≡ s2 d1

(mod gcd(d1 d2 , N ))

52

3. Modular Forms of Weight 2

where sj satisfies cj sj ≡ 1 (mod dj ).

In SAGE the command boundary map() computes the boundary map from M2 (Γ0 (N )) to B2 (Γ0 (N )), and the cuspidal submodule() command computes its kernel. For example, for level 2 the boundary map is given by the matrix [1 − 1], and its kernel is the 0 space: sage: M = ModularSymbols(2, 2) sage: M.boundary_map() Hecke module morphism boundary map defined by the matrix [ 1 -1] Domain: Modular Symbols space of dimension 1 for Gamma_0(2) of weight ... Codomain: Space of Boundary Modular Symbols for Congruence Subgroup Gamma0(2) ... sage: M.cuspidal_submodule() Modular Symbols subspace of dimension 0 of Modular Symbols space of dimension 1 for Gamma_0(2) of weight 2 with sign 0 over Rational Field The smallest level for which the boundary map has nontrivial kernel, i.e., for which S2 (Γ0 (N )) 6= 0, is N = 11. sage: M = ModularSymbols(11, 2) sage: M.boundary_map().matrix() [ 1 -1] [ 0 0] [ 0 0] sage: M.cuspidal_submodule() Modular Symbols subspace of dimension 2 of Modular Symbols space of dimension 3 for Gamma_0(11) of weight 2 with sign 0 over Rational Field sage: S = M.cuspidal_submodule(); S Modular Symbols subspace of dimension 2 of Modular Symbols space of dimension 3 for Gamma_0(11) of weight 2 with sign 0 over Rational Field sage: S.basis() ((1,8), (1,9)) The following illustrates that the Hecke operators preserve S2 (Γ0 (N )):

3.6. Computing a Basis for S2 (Γ0 (N ))

53

sage: S.T(2).matrix() [-2 0] [ 0 -2] sage: S.T(3).matrix() [-1 0] [ 0 -1] sage: S.T(5).matrix() [1 0] [0 1] A nontrivial fact is that for p prime the eigenvalue of each of these matrices is p + 1 − #E(Fp ), where E is the elliptic curve X0 (11) defined by the (affine) equation y 2 + y = x3 − x2 − 10x − 20. For example, we have sage: sage: -2 sage: -1 sage: 1 sage: -2

E = EllipticCurve([0,-1,1,-10,-20]) 2 + 1 - E.Np(2) 3 + 1 - E.Np(3) 5 + 1 - E.Np(5) 7 + 1 - E.Np(7)

The same numbers appear as the eigenvalues of Hecke operators: sage: [S.T(p).matrix()[0,0] for p in [2,3,5,7]] [-2, -1, 1, -2] In fact, something similar happens for every elliptic curve over Q. The book [DS05] (especially Chapter 8) is about this striking numerical relationship between the number of points on elliptic curves modulo p and coefficients of modular forms.

3.6. Computing a Basis for S2 (Γ0 (N )) This section is about a method for using modular symbols to compute a basis for S2 (Γ0 (N )). It is not the most efficient for certain applications, but it is easy to explain and understand. See Section 3.7 for a method that takes advantage of additional structure of S2 (Γ0 (N )).

54

3. Modular Forms of Weight 2

Let M2 (Γ0 (N ); Q) and S2 (Γ0 (N ); Q) be the spaces of modular symbols and cuspidal modular symbols over Q. Before we begin, we describe a simple but crucial fact about the relation between cusp forms and Hecke operators. P If f = bn q n ∈ C[[q]] is a power series, let an (f ) = bn be the n coefficient of f . Notice that an is a C-linear map C[[q]] → C. As explained in [DS05, Prop. 5.3.1] and [Lan95, §VII.3] (recall also Proposition 2.31), the Hecke operators Tn act on elements of M2 (Γ0 (N )) as follows (where k = 2 below):   ! Ã ∞ ∞ X X X  ε(d) · dk−1 · amn/d2  q m , am q m = (3.6.1) Tn m=0

m=0

1≤d | gcd(n,m)

where ε(d) = 1 if gcd(d, N ) = 1 and ε(d) = 0 if gcd(d, N ) 6= 1. (Note: More generally, if f ∈ Mk (Γ1 (N )) is a modular form with Dirichlet character ε, then the above formula holds; above we are considering this formula in the special case when ε is the trivial character and k = 2.) Lemma 3.22. Suppose f ∈ C[[q]] and n is a positive integer. Let Tn be the operator on q-expansions (formal power series) defined by (3.6.1). Then a1 (Tn (f )) = an (f ). Proof. The coefficient of q in (3.6.1) is ε(1) · 1 · a1·n/12 = an .

¤

The Hecke algebra T is the ring generated by all Hecke operators Tn acting on Mk (Γ1 (N )). Let T′ denote the image of the Hecke algebra in End(S2 (Γ0 (N ))), and let T′C = T′ ⊗Z C be the C-span of the Hecke operators. ˜ C denote the subring of End(C[[q]]) generated over C by all Hecke Let T operators acting on formal power series via definition (3.6.1). Proposition 3.23. There is a bilinear pairing of complex vector spaces ˜C → C C[[q]] × T given by hf, ti = a1 (t(f )). ˜ C , then f = 0. If f is such that hf, ti = 0 for all t ∈ T Proof. The pairing is bilinear since both t and a1 are linear. ˜ C . Then hf, Tn i = 0 Suppose f ∈ C[[q]] is such that hf, ti = 0 for all t ∈ T for each positive integer n. But by Lemma 3.22 we have an (f ) = a1 (Tn (f )) = 0 for all n; thus f = 0.

¤

3.6. Computing a Basis for S2 (Γ0 (N ))

55

Proposition 3.24. There is a perfect bilinear pairing of complex vector spaces S2 (Γ0 (N )) × T′C → C given by hf, ti = a1 (t(f )). Proof. The pairing has 0 kernel on the left by Proposition 3.23. Suppose that t ∈ T′C is such that hf, ti = 0 for all f ∈ S2 (Γ0 (N )). Then a1 (t(f )) = 0 for all f . For any n, the image Tn (f ) is also a cusp form, so a1 (t(Tn (f ))) = 0 for all n and f . Finally the fact that T′ is commutative and Lemma 3.22 together imply that for all n and f , 0 = a1 (t(Tn (f ))) = a1 (Tn (t(f ))) = an (t(f )), so t(f ) = 0 for all f . Thus t is the 0 operator. Since S2 (Γ0 (N )) has finite dimension and the kernel on each side of the pairing is 0, it follows that the pairing is perfect, i.e., defines an isomorphism T′ ∼ = HomC (S2 (Γ0 (N )); C). C

¤ By Proposition 3.24 there is an isomorphism of vector spaces (3.6.2)

∼ =

Ψ : S2 (Γ0 (N )) −−→ Hom(T′C , C)

that sends f ∈ S2 (Γ0 (N )) to the homomorphism t 7→ a1 (t(f )).

For any C-linear map ϕ : T′C → C, let fϕ =

∞ X

n=1

ϕ(Tn )q n ∈ C[[q]].

Lemma 3.25. The series fϕ is the q-expansion of Ψ−1 (ϕ) ∈ S2 (Γ0 (N )). Proof. Note that it is not even a priori obvious that fϕ is the q-expansion of a modular form. Let g = Ψ−1 (ϕ), which is by definition the unique element of S2 (Γ0 (N )) such that hg, Tn i = ϕ(Tn ) for all n. By Lemma 3.22, we have hfϕ , Tn i = a1 (Tn (fϕ )) = an (fϕ ) = ϕ(Tn ),

so hfϕ − g, Tn i = 0 for all n. Proposition 3.23 implies that fϕ − g = 0, so fϕ = g = Ψ−1 (ϕ), as claimed. ¤ Conclusion: The cusp forms fϕ , as ϕ varies through a basis of Hom(T′C , C), form a basis for S2 (Γ0 (N )). In particular, we can compute S2 (Γ0 (N )) by computing Hom(T′C , C), where we compute T′ in any way we want, e.g., using a space that contains an isomorphic copy of S2 (Γ0 (N )).

56

3. Modular Forms of Weight 2

Algorithm 3.26 (Basis of Cusp Forms). Given positive integers N and B, this algorithm computes a basis for S2 (Γ0 (N )) to precision O(q B ). (1) Compute M2 (Γ0 (N ); Q) via the presentation of Section 3.3.2. (2) Compute the subspace S2 (Γ0 (N ); Q) of cuspidal modular symbols as in Section 3.5. (3) Let d = 12 ·dim S2 (Γ0 (N ); Q). By Proposition 3.8, d is the dimension of S2 (Γ0 (N )). (4) Let [Tn ] denote the matrix of Tn acting on a basis of S2 (Γ0 (N ); Q). For a matrix A, let aij (A) denote the ijth entry of A. For various integers i, j with 0 ≤ i, j ≤ d − 1, compute formal q-expansions fij (q) =

B−1 X n=1

aij ([Tn ])q n + O(q B ) ∈ Q[[q]]

until we find enough to span a space of dimension d (or exhaust all of them). These fij are a basis for S2 (Γ0 (N )) to precision O(q B ). 3.6.1. Examples. We use SAGE to demonstrate Algorithm 3.26. Example 3.27. The smallest N with S2 (Γ0 (N )) 6= 0 is N = 11.

sage: M = ModularSymbols(11); M.basis() ((1,0), (1,8), (1,9)) sage: S = M.cuspidal_submodule(); S Modular Symbols subspace of dimension 2 of Modular Symbols space of dimension 3 for Gamma_0(11) of weight 2 with sign 0 over Rational Field

We compute a few Hecke operators, and then read off a nonzero cusp form, which forms a basis for S2 (Γ0 (11)): sage: S.T(2).matrix() [-2 0] [ 0 -2] sage: S.T(3).matrix() [-1 0] [ 0 -1] Thus f0,0 = q − 2q 2 − q 3 + · · · ∈ S2 (Γ0 (11)) forms a basis for S2 (Γ0 (11)). Example 3.28. We compute a basis for S2 (Γ0 (33)) to precision O(q 6 ).

3.6. Computing a Basis for S2 (Γ0 (N ))

57

sage: M = ModularSymbols(33) sage: S = M.cuspidal_submodule(); S Modular Symbols subspace of dimension 6 of Modular Symbols space of dimension 9 for Gamma_0(33) of weight 2 with sign 0 over Rational Field Thus dim S2 (Γ0 (33)) = 3. sage: R. = PowerSeriesRing(QQ) sage: v = [S.T(n).matrix()[0,0] for n in range(1,6)] sage: f00 = sum(v[n-1]*q^n for n in range(1,6)) + O(q^6) sage: f00 q - q^2 - q^3 + q^4 + O(q^6) This gives us one basis element of S2 (Γ0 (33)). It remains to find two others. We find sage: v = [S.T(n).matrix()[0,1] for n in range(1,6)] sage: f01 = sum(v[n-1]*q^n for n in range(1,6)) + O(q^6) sage: f01 -2*q^3 + O(q^6) and sage: sage: sage: q^3 +

v = [S.T(n).matrix()[1,0] for n in range(1,6)] f10 = sum(v[n-1]*q^n for n in range(1,6)) + O(q^6) f10 O(q^6)

This third one is (to our precision) a scalar multiple of the second, so we look further. sage: v = [S.T(n).matrix()[1,1] for n in range(1,6)] sage: f11 = sum(v[n-1]*q^n for n in range(1,6)) + O(q^6) sage: f11 q - 2*q^2 + 2*q^4 + q^5 + O(q^6) This latter form is clearly not in the span of the first two. Thus we have the following basis for S2 (Γ0 (33)) (to precision O(q 6 )):

f00 = q − q 2 − q 3 + q 4 + · · · ,

f11 = q − 2q 2 + 2q 4 + q 5 + · · · , f10 = q 3 + · · · .

58

3. Modular Forms of Weight 2

Example 3.29. Next consider N = 23, where we have d = dim S2 (Γ0 (23)) = 2. The command q expansion cuspforms computes matrices Tn and returns a function f such that f (i, j) is the q-expansion of fi,j to some precision. (For efficiency reasons, f (i, j) in SAGE actually computes matrices of Tn acting on a basis for the linear dual of S2 (Γ0 (N )).) sage: M = ModularSymbols(23) sage: S = M.cuspidal_submodule() sage: S Modular Symbols subspace of dimension 4 of Modular Symbols space of dimension 5 for Gamma_0(23) of weight 2 with sign 0 over Rational Field sage: f = S.q_expansion_cuspforms(6) sage: f(0,0) q - 2/3*q^2 + 1/3*q^3 - 1/3*q^4 - 4/3*q^5 + O(q^6) sage: f(0,1) O(q^6) sage: f(1,0) -1/3*q^2 + 2/3*q^3 + 1/3*q^4 - 2/3*q^5 + O(q^6) Thus a basis for S2 (Γ0 (23)) is 2 1 1 4 f0,0 = q − q 2 + q 3 − q 4 − q 5 + · · · , 3 3 3 3 1 2 2 3 1 4 2 5 f1,0 = − q + q + q − q + · · · . 3 3 3 3 Or, in echelon form, q − q3 − q4 + · · ·

q 2 − 2q 3 − q 4 + 2q 5 + · · ·

which we computed using sage: S.q_expansion_basis(6) [ q - q^3 - q^4 + O(q^6), q^2 - 2*q^3 - q^4 + 2*q^5 + O(q^6) ]

3.7. Computing S2 (Γ0 (N )) Using Eigenvectors In this section we describe how to use modular symbols to construct a basis of S2 (Γ0 (N )) consisting of modular forms that are eigenvectors for every

3.7. Computing S2 (Γ0 (N )) Using Eigenvectors

59

element of the ring T(N ) generated by the Hecke operator Tp , with p ∤ N . Such eigenvectors are called eigenforms. Suppose M is a positive integer that divides N . As explained in [Lan95, VIII.1–2], for each divisor d of N/M there is a natural degeneracy map αM,d : S2 (Γ0 (M )) → S2 (Γ0 (N )) given by αM,d (f (q)) = f (q d ). The new subspace of S2 (Γ0 (N )), denoted S2 (Γ0 (N ))new , is the complementary T-submodule of the T-module generated by the images of all maps αM,d , with M and d as above. It is a nontrivial fact that this complement is well defined; one possible proof uses the Petersson inner product (see [Lan95, §VII.5]). The theory of Atkin and Lehner [AL70] (see Theorem 9.4 below) asserts that, as a T(N ) -module, S2 (Γ0 (N )) decomposes as follows: S2 (Γ0 (N ))

=

M

βM,d (S2 (Γ0 (M ))new ).

M |N, d|N/M

To compute S2 (Γ0 (N )) it suffices to compute S2 (Γ0 (M ))new for each M | N .

We now turn to the problem of computing S2 (Γ0 (N ))new . Atkin and Lehner [AL70] proved that S2 (Γ0 (N ))new is spanned by eigenforms for all Tp with p ∤ N and that the common eigenspaces of all the Tp with p ∤ N each have dimension 1. Moreover, if f ∈ S2 (Γ0 (N ))new is an eigenform then the coefficient of q in the q-expansion of f is nonzero, so it is possible to normalize f so the coefficient of q is 1 (such a normalized eigenform in the new subspace is called a newform). With f so normalized, if Tp (f ) = ap f , P∞ n then the pth Fourier coefficient of f is ap . If f = n=1 an q is a normalized eigenvector for all Tp , then the an , with n composite, are determined by the ap , with p prime, by the following formulas: anm = an am when n and m are relatively prime and apr = apr−1 ap − papr−2 for p ∤ N prime. When p | N , apr = arp . We conclude that in order to compute S2 (Γ0 (N ))new , it suffices to compute all systems of eigenvalues {a2 , a3 , a5 , . . .} of the prime-indexed Hecke operators T2 , T3 , T5 , . . . acting on S2 (Γ0 (NP ))new . Given a system of n eigenvalues, the corresponding eigenform is f = ∞ n=1 an q , where the an , for n composite, are determined by the recurrence given above.

In light of the pairing h , i introduced in Section 3.1, computing the above systems of eigenvalues {a2 , a3 , a5 , . . .} amounts to computing the systems of eigenvalues of the Hecke operators Tp on the subspace V of S2 (Γ0 (N )) that corresponds to the new subspace of S2 (Γ0 (N )). For each proper divisor M of N and each divisor ¡ d¢ of N/M , let φM,d : S2 (Γ0 (N )) → S2 (Γ0 (M )) be the map sending x to d0 10 x. Then V is the intersection of the kernels of all maps φM,d . Computing the systems of eigenvalues of a collection of commuting diagonalizable endomorphisms is a problem in linear algebra (see Chapter 7).

60

3. Modular Forms of Weight 2

Example 3.30. All forms in S2 (Γ0 (39)) are new. Up to Galois conjugacy, the eigenvalues of the Hecke operators T2 , T3 , T5 , and T7 on S2 (Γ0 (39)) are {1, −1, 2, −4} and {a, 1, −2a − 2, 2a + 2}, where a2 + 2a − 1 = 0. Each of these eigenvalues occur in S2 (Γ0 (39)) with multiplicity two; for example, the characteristic polynomial of T2 on S2 (Γ0 (39)) is (x − 1)2 · (x2 + 2x − 1)2 . Thus S2 (Γ0 (39)) is spanned by f1 = q + q 2 − q 3 − q 4 + 2q 5 − q 6 − 4q 7 + · · · ,

f2 = q + aq 2 + q 3 + (−2a − 1)q 4 + (−2a − 2)q 5 + aq 6 + (2a + 2)q 7 + · · · , f3 = q + σ(a)q 2 + q 3 + (−2σ(a) − 1)q 4 + (−2σ(a) − 2)q 5 + σ(a)q 6 + · · · ,

where σ(a) is the other Gal(Q/Q)-conjugate of a. 3.7.1. Summary. To compute the q-expansion of a basis for S2 (Γ0 (N )), we use the degeneracy maps so that we only have to solve the problem for S2 (Γ0 (M ))new , for all integers M | N . Using modular symbols, we compute all systems of eigenvalues {a2 , a3 , a5 , . . .}, and then write down the P corresponding eigenforms an q n .

3.8. Exercises

3.1 Suppose that λ, λ′ ∈ h are in the same orbit for the action of Γ0 (N ), i.e., that there exists g ∈ Γ0 (N ) such that g(λ) = λ′ . Let Λ = Z + Zλ and Λ′ = Z + Zλ′ . Prove that the pairs (C/Λ, ( N1 Z + Λ)/Λ) and (C/Λ′ , ( N1 Z + Λ′ )/Λ′ ) are isomorphic. (By an isomorphism (E, C) → (F, D) of pairs, we mean an isomorphism φ : E → F of elliptic curves that sends C to D. You may use the fact that an isomorphism of elliptic curves over C is a C-linear map C → C that sends the lattice corresponding to one curve onto the lattice corresponding to the other.) 3.2 Let n, m be integers and N a positive integer. Prove that the modular symbol {n, m} is 0 as an element of M2 (Γ0 (N )). [Hint: See Example 3.6.] 3.3 Let p be a prime. (a) List representative elements of P1 (Z/pZ). (b) What is the cardinality of P1 (Z/pZ) as a function of p? (c) Prove that there is a bijection between the right cosets¡of Γ¢0 (p) in SL2 (Z) and the elements of P1 (Z/pZ) that sends ac db to (c : d). (As mentioned in this chapter, the analogous statement is also true when the level is composite; see [Cre97a, §2.2] for complete details.) 3.4 Use the inductive proof of Proposition 3.11 to write {0, 4/7} in terms of Manin symbols for Γ0 (7).

3.8. Exercises

61

3.5 Show that the Hecke operator T2 acts as multiplication by 3 on the space M2 (Γ0 (3)) as follows: (a) Write down right coset representatives for Γ0 (3) in SL2 (Z). (b) List all eight relations coming from Theorem 3.13. (c) Find a single Manin symbols [ri ] so that the three other Manin symbols are a nonzero multiple of [ri ] modulo the relations found in the previous step. (d) Use formula (3.4.1) to compute T2 ([ri ]). You will obtain a sum of four symbols. Using the relations above, write this sum as a multiple of [ri ]. (The multiple must be 3 or you made a mistake.)

Chapter 4

Dirichlet Characters

In this chapter we develop a theory for computing with Dirichlet characters, which are extremely important to computations with modular forms for (at least) two reasons: (1) To compute the Eisenstein subspace Ek (Γ1 (N )) of Mk (Γ1 (N )), we write down Eisenstein series attached to pairs of Dirichlet characters (the space Ek (Γ1 (N )) will be defined in Chapter 5). (2) To compute Sk (Γ1 (N )), we instead compute a decomposition M Mk (Γ1 (N )) = Mk (Γ1 (N ), ε)

and then compute each factor (see Section 9.1). Here the sum is over all Dirichlet characters ε of modulus N .

Dirichlet characters appear frequently in many other areas of number theory. For example, by the Kronecker-Weber theorem, Dirichlet characters correspond to the 1-dimensional representations of Gal(Q/Q). After defining Dirichlet characters in Section 4.1, in Section 4.2 we describe a good way to represent Dirichlet characters using a computer. Section 4.3 is about how to evaluate Dirichlet characters and leads naturally to a discussion of the baby-step giant-step algorithm for solving the discrete log problem and methods for efficiently computing the Kronecker symbol. In Section 4.4 we explain how to factor Dirichlet characters into their prime power constituents and apply this to the computations of conductors. We describe how to carry out a number of standard operations with Dirichlet characters in Section 4.6 and discuss alternative ways to represent them in Section 4.7. Finally, in Section 4.8 we give a very short tutorial about how to compute with Dirichlet characters using SAGE. 63

64

4. Dirichlet Characters

4.1. The Definition Fix an integral domain R and a root ζ of unity in R. Definition 4.1 (Dirichlet Character). A Dirichlet character of modulus N over R is a map ε : Z → R such that there is a homomorphism f : (Z/N Z)∗ → hζi for which ( 0 if gcd(a, N ) > 1, ε(a) = f (a mod N ) if gcd(a, N ) = 1. We denote the group of such Dirichlet characters by D(N, R). Note that elements of D(N, R) are in bijection with homomorphisms (Z/N Z)∗ → hζi. ³ ´ A familiar Dirichlet character is the Legendre symbol ap , with p an odd prime, that appears in quadratic reciprocity theory. It is a Dirichlet character of modulus p that takes the value 1 on integers that are congruent to a nonzero square modulo p, the value −1 on integers that are congruent to a nonzero nonsquare modulo p, and 0 on integers divisible by p.

4.2. Representing Dirichlet Characters Lemma 4.2. The groups (Z/N Z)∗ and D(N, C) are isomorphic. Proof. We prove the more general fact that for any finite abelian group G, we have that G ≈ Hom(G, C∗ ). To deduce this latter isomorphism, first reduce to the case when G is cyclic by writing G as a product of cyclic groups. The cyclic case follows because if G is cyclic of order n, then C∗ contains an nth root of unity, so Hom(G, C∗ ) is also cyclic of order n. Any two cyclic groups of the same order are isomorphic, so G and Hom(G, C∗ ) are isomorphic. ¤ Corollary 4.3. We have #D(N, R) | ϕ(N ), with equality if and only if the order of our choice of ζ ∈ R is a multiple of the exponent of the group (Z/N Z)∗ . Proof. This is because #(Z/N Z)∗ = ϕ(N ).

¤

Fix a positive integer N . To find Q the set of “canonical” generators for the group (Z/N Z)∗ , write N = ni=0 pei i where p0 < p1 < · · · < pn are the prime divisors of N . By Exercise 4.2, each factor (Z/pei i Z)∗ is a cyclic group Ci = hgi i, except if p0 = 2 and e0 ≥ 3, in which case (Z/pe00 Z)∗ is a product of the cyclic subgroup C0 = h−1i of order 2 with the cyclic subgroup C1 = h5i. In all cases we have Y Y (Z/N Z)∗ ∼ Ci = hgi i. = 0≤i≤n

0≤i≤n

65

4.2. Representing Dirichlet Characters

For i such that pi > 2, choose the generator gi of Ci to be the element of {2, 3, . . . , pei i − 1} that is smallest and generates. Finally, use the Chinese Remainder Theorem (see [Coh93, §1.3.3]) to lift each gi to an element in e (Z/N Z)∗ , also denoted gi , that is 1 modulo each pj j for j 6= i. Algorithm 4.4 (Minimal Generator for (Z/pr Z)∗ ). Given a prime power pr with p odd, this algorithm computes the minimal generator of (Z/pr Z)∗ . r r−1 · 2 · ((p − 1)/2) as a (1) [Factor Group Q ei Order] Factor n = φ(p ) = p product pi of primes. This is equivalent in difficulty to factoring (p − 1)/2. (See, e.g., [Coh93, Ch.8, Ch. 10] for an excellent discussion of factorization algorithms, though of course much progress has been made since then.) (2) [Initialize] Set g = 2. (3) [Generator?] Using the binary powering algorithm (see [Coh93, §1.2]), compute g n/pi (mod pr ), for each prime divisor pi of n. If any of these powers are 1, then g is not a generator, so set g = g + 1 and go to step (2). If no powers are 1, output g and terminate.

See Exercise 4.3 for a proof that this algorithm is correct. Example 4.5. A minimal generator for (Z/49Z)∗ is 3. We have n = ϕ(49) = 42 = 2 · 3 · 7 and 2n/2 ≡ 1,

2n/3 ≡ 18,

2n/7 ≡ 15 (mod 49),

so 2 is not a generator for (Z/49Z)∗ . (We see this just from 2n/2 ≡ 1 (mod 49).) However 3 is a generator since 3n/2 ≡ 48,

3n/3 ≡ 30,

3n/7 ≡ 43 (mod 49).

Example 4.6. In this example we compute minimal generators for N = 25, 100, and 200: (1) The minimal generator for (Z/25Z)∗ is 2. (2) The minimal generators for (Z/100Z)∗ , lifted to numbers modulo 100, are g0 = 51 and g1 = 77. Notice that g0 ≡ −1 (mod 4) and g0 ≡ 1 (mod 25) and that g1 ≡ 2 (mod 25) is the minimal generator modulo 25. (3) The minimal generators for (Z/200Z)∗ , lifted to numbers modulo 200, are g0 = 151, g1 = 101, and g2 = 177. Note that g0 ≡ −1 (mod 4), that g1 ≡ 5 (mod 8) and g2 ≡ 2 (mod 25). In SAGE, the command Integers(N) creates Z/N Z.

66

4. Dirichlet Characters

sage: R = Integers(49) sage: R Ring of integers modulo 49 The unit gens command computes the minimal generators for (Z/N Z)∗ , as defined above. sage: R.unit_gens() [3] sage: Integers(25).unit_gens() [2] sage: Integers(100).unit_gens() [51, 77] sage: Integers(200).unit_gens() [151, 101, 177] sage: Integers(2005).unit_gens() [402, 1206] sage: Integers(200000000).unit_gens() [174218751, 51562501, 187109377] Fix an element ζ of finite multiplicative order in a ring R, and let D(N, R) denote the group of Dirichlet characters of modulus N over R, with image in hζi ∪ {0}. In most of this chapter, we specify an element ε ∈ D(N, R) by giving the list (4.2.1)

[ε(g0 ), ε(g1 ), . . . , ε(gn )]

of images of the generators of (Z/N Z)∗ . (Note that if N is even, the number of elements of the list (4.2.1) does depend on whether or not 8 | N —there are two factors corresponding to 2 if 8 | N , but only one if 8 ∤ N .) This representation completely determines ε and is convenient for arithmetic operations. It is analogous to representing a linear transformation by a matrix. Remark 4.7. In any actual implementation (e.g., the one in SAGE), it is better to represent the ε(gi ) by recording an integer j such that ε(gi ) = ζ j , where ζ ∈ R is a fixed root of unity. Then (4.2.1) is internally represented as an element of (Z/mZ)n+1 , where m is the multiplicative order of ζ. When the representation of (4.2.1) is needed for an algorithm, it can be quickly computed on the fly using a table of the powers of ζ. See Section 4.7 for further discussion about ways to represent characters. Example 4.8. The group D(5, C) has elements {[1], [i], [−1], [−i]}, so it is cyclic of order ϕ(5) = 4. In contrast, the group D(5, Q) has only the two

67

4.3. Evaluation of Dirichlet Characters

elements [1] and [−1] and order 2. The command DirichletGroup(N) with no second argument creates the group of Dirichlet characters with values in the cyclotomic field Q(ζn ), where n is the exponent of the group (Z/N Z)∗ . Every element in D(N, C) takes values in Q(ζn ), so D(N, Q(ζn )) ≈ D(N, C). sage: [[1], sage: [[1],

list(DirichletGroup(5)) [zeta4], [-1], [-zeta4]] list(DirichletGroup(5, QQ)) [-1]]

4.3. Evaluation of Dirichlet Characters This section is about how to compute ε(n), where ε is a Dirichlet character and n is an integer. We begin with an example. Example 4.9. If N = 200, then g0 = 151, g1 = 101 and g2 = 177, as we saw in Example 4.6. The exponent of (Z/200Z)∗ is 20, since that is the least common multiple of the exponents of 4 = #(Z/8Z)∗ and 20 = #(Z/25Z)∗ . The orders of g0 , g1 , and g2 are 2, 2, and 20. Let ζ = ζ20 be a primitive 20th root of unity in C. Then the following are generators for D(200, C): ε0 = [−1, 1, 1],

ε1 = [1, −1, 1],

ε2 = [1, 1, ζ],

and ε = [1, −1, ζ 5 ] is an example element of order 4. To evaluate ε(3), we write 3 in terms of g0 , g1 , and g2 . First, reducing 3 modulo 8, we see that 3 ≡ g0 ·g1 (mod 8). Next reducing 3 modulo 25 and trying powers of g2 = 2, we find that e ≡ g27 (mod 25). Thus ε(3) = ε(g0 · g1 · g27 )

= ε(g0 )ε(g1 )ε(g2 )7

= 1 · (−1) · (ζ 5 )7

= −ζ 35 = −ζ 15 .

We next illustrate the above computation of ε(3) in SAGE. First we make the group D(200, Q(ζ8 )) and list its generators.

68

4. Dirichlet Characters

sage: G = DirichletGroup(200) sage: G Group of Dirichlet characters of modulus 200 over Cyclotomic Field of order 20 and degree 8 sage: G.exponent() 20 sage: G.gens() ([-1, 1, 1], [1, -1, 1], [1, 1, zeta20]) We construct ε. sage: K = G.base_ring() sage: zeta = K.0 sage: eps = G([1,-1,zeta^5]) sage: eps [1, -1, zeta20^5] Finally, we evaluate ε at 3. sage: eps(3) zeta20^5 sage: -zeta^15 zeta20^5 Example 4.9 illustrates that if ε is represented using a list as described above, evaluation of ε is inefficient without extra information; it requires solving the discrete log problem in (Z/N Z)∗ . Remark 4.10. For a general character ε, is calculation of ε at least as hard as finding discrete logarithms? Quadratic characters are easier—see Algorithm 4.23. Algorithm 4.11 (Evaluate ε). Given a Dirichlet character ε of modulus N , represented by a list [ε(g0 ), ε(g1 ), . . . , ε(gn )], and an integer a, this algorithm computes ε(a). (1) [GCD] Compute g = gcd(a, N ). If g > 1, output 0 and terminate. (2) [Discrete Log] For each i, write a (mod pei i ) as a power mi of gi using some algorithm for solving the discrete log problem (see below). If pi = 2, write a (mod pei i ) as (−1)m0 · 5m1 . (This step is analogous to writing a vector in terms of a basis.)

69

4.3. Evaluation of Dirichlet Characters

Q (3) [Multiply] Output ε(gi )mi as an element of R, and terminate. (This is analogous to multiplying a matrix times a vector.) 4.3.1. The Discrete Log Problem. Exercise 4.4 gives an isomorphism of groups (1 + pn−1 (Z/pn Z), ×) ∼ = (Z/pZ, +), so one sees by induction that step (2) is “about as difficult” as finding a discrete log in (Z/pZ)∗ . There is an algorithm called “baby-step giant√ step”, which solves the discrete log problem in (Z/pZ)∗ in time O( ℓ), where ℓ is the largest prime factor of p − 1 = #(Z/pZ)∗ (note that the discrete log problem in (Z/pZ)∗ reduces to a series of discrete log problems in each prime-order cyclic factor). This is √ unfortunately still exponential in the number of digits of ℓ; it also uses O( ℓ) memory. We now describe this algorithm without any specific optimizations. Algorithm 4.12 (Baby-step Giant-step Discrete Log). Given a prime p, a generator g of (Z/pZ)∗ , and an element a ∈ (Z/pZ)∗ , this algorithm finds an n such that g n = a. (Note that this algorithm works in any cyclic group, not just (Z/pZ)∗ .) √ √ (1) [Make Lists] Let m = ⌈ p⌉ be the ceiling of p, and construct two lists 1, g m , . . . , g (m−1)m

(giant steps)

and a, ag, ag 2 , . . . , ag m−1

(baby steps).

(2) [Find Match] Sort the two lists and find a match g im = ag j . Then a = g im−j . Proof. We prove that there will always be a match. Since we know that a = g k for some k with 0 ≤ k ≤ p − 1 and any such k can be written in the form im − j for 0 ≤ i, j ≤ m − 1, we will find such a match. ¤ Algorithm 4.12 uses nothing special about (Z/pZ)∗ , so it works in a generic group. It is a theorem that there is no faster algorithm to find discrete logs in a “generic group” (see [Sho97, Nec94]). There are much better subexponential algorithms for solving the discrete log problem in (Z/pZ)∗ , which use the special structure of this group. They use the number field sieve (see, e.g., [Gor93]), which is also the best-known algorithm for factoring integers. This class of algorithms has been very well studied by cryptographers; though sub-exponential, solving discrete log problems when p is large is still extremely difficult. For a more in-depth survey see [Gor04]. For computing Dirichlet characters in our context, p is not too large, so Algorithm 4.12 works well.

70

4. Dirichlet Characters

4.3.2. Enumeration of All Values. For many applications of Dirichlet characters to computing modular forms, N is fairly small, e.g., N < 106 , and we evaluate ε on a huge number of random elements, inside inner loops of algorithms. Thus for such purposes it will often be better to make a table of all values of ε, so that evaluation of ε is extremely fast. The following algorithm computes a table of all values of ε, and it does not require computing any discrete logs since we are computing all values. Algorithm 4.13 (Values of ε). Given a Dirichlet character ε represented by the list of values of ε on the minimal generators gi of (Z/N Z)∗ , this algorithm creates a list of all the values of ε. Q (1) [Initialize] For each minimal generator gi , set ai = 0. Let n = giai , and set z = 1. Create a list v of N values, all initially set equal to 0. When this algorithm terminates, the list v will have the property that v [x (mod N )] = ε(x). Notice that we index v starting at 0. (2) [Add Value to Table] Set v[n] = z. (3) [Finished?] If each ai is one less than the order of gi , output v and terminate. (4) [Increment] Set a0 = a0 + 1, n = n · g0 (mod N ), and z = z · ε(g0 ). If a0 ≥ ord(g0 ), set a0 → 0, and then set a1 = a1 + 1, n = n · g1 (mod N ), and z = z · ε(g1 ). If a1 ≥ ord(g1 ), do what you just did with a0 but with all subscripts replaced by 1. Etc. (Imagine a car odometer.) Go to step (2).

4.4. Conductors of Dirichlet Characters The following algorithm for computing the order of ε reduces the problem to computing the orders of powers of ζ in R. Algorithm 4.14 (Order of Character). This algorithm computes the order of a Dirichlet character ε ∈ D(N, R). (1) Compute the order ri of each ε(gi ), for each minimal generator gi of (Z/N Z)∗ . The order of ε(gi ) is a divisor of n = #(Z/pei i Z)∗ so we can compute its order by considering the divisors of n. (2) Compute and output the least common multiple of the integers ri .

Remark 4.15. Computing the order of ε(gi ) ∈ R is potentially difficult. Simultaneously using a different representation of Dirichlet characters avoids having to compute the order of elements of R (see Section 4.7). The next algorithm factors ε as a product of “local” characters, one for each prime divisor of N . It is useful for other algorithms, e.g., for explicit

4.4. Conductors of Dirichlet Characters

71

computations with trace formulas (see [Hij74]). This factorization is easy to compute because of how we represent ε. Algorithm 4.16 (Factorization Q ei of Character). Given a Dirichlet character ε ∈ D(N, R), with N = pi , this algorithm finds Dirichlet Q characters εi modulo pei i , such that for all a ∈ (Z/N Z)∗ , we have ε(a) = εi (a (mod pei i )). If 2 | N , the steps are as follows: (1) Let gi be the minimal generators of (Z/N Z)∗ , so ε is given by a list [ε(g0 ), . . . , ε(gn )]. (2) For i = 2, . . . , n, let εi be the element of D(pei i , R) defined by the singleton list [ε(gi )]. (3) Let ε1 be the element of D(2e1 , R) defined by the list [ε(g0 ), ε(g1 )] of length 2. Output the εi and terminate. If 2 ∤ N , then omit step (3), and include all i in step (2). The factorization of Algorithm 4.16 is unique since each εi is determined by the image of the canonical map (Z/pei i Z)∗ in (Z/N Z)∗ , which sends a e (mod pei i ) to the element of (Z/N Z)∗ that is a (mod pei i ) and 1 (mod pj j ) for j 6= i.

Example 4.17. If ε = [1, −1, ζ 5 ] ∈ D(200, C), then ε1 = [1, −1] ∈ D(8, C) and ε2 = [ζ 5 ] ∈ D(25, C).

Definition 4.18 (Conductor). The conductor of a Dirichlet character ε ∈ D(N, R) is the smallest positive divisor c | N such that there is a character ε′ ∈ D(c, R) for which ε(a) = ε′ (a) for all a ∈ Z with (a, N ) = 1. A Dirichlet character is primitive if its modulus equals its conductor. The character ε′ associated to ε with modulus equal to the conductor of ε is called the primitive character associated to ε. We will be interested in conductors later, when computing new subspaces of spaces of modular forms with character. Also certain formulas for special values of L functions are only valid for primitive characters. Algorithm 4.19 (Conductor). This algorithm computes the conductor of a Dirichlet character ε ∈ D(N, R). (1) [Factor Character] Using Algorithm 4.16, find characters εi whose product is ε. (2) [Compute Orders] Using Algorithm 4.14, compute the orders ri of each εi . (3) [Conductors of Factors] For each i, either set ci → 1 if εi is the ordp (r )+1 trivial character (i.e., of order 1) or set ci = pi i i , where ordp (n) is the largest power of p that divides n.

72

4. Dirichlet Characters

(4) [Adjust at 2?] If p1 = 2 and ε1 (5) 6= 1, set c1 = 2c1 . Q (5) [Finished] Output c = ci and terminate. Proof. Let εi be the local factors of ε, as in step (1). We first show that the product of the conductors fi of the εi is the conductor f of ε. Since Q εi factors ∗ ∗ through (Z/fi Z) , the product Q ε of the εi factors through (Z/ fi Z) , so the conductor of ε divides fi . Conversely, if ordpi (f ) < ordpi (fi ) for some i, then we could factor ε as a product of local (prime power) characters differently, which contradicts that this factorization is unique. It remains to prove that if ε is a nontrivial character of modulus pn , where p is a prime, and if r is the order of ε, then the conductor of ε is pordp (r)+1 , except possibly if 8 | pn . Since the order and conductor of ε and of the associated primitive character ε′ are the same, we may assume ε is primitive, i.e., that pn is the conductor of ε; note that n > 0, since ε is nontrivial. First suppose p is odd. Then the abelian group D(pn , R) splits as a direct sum D(p, R) ⊕ D(pn , R)′ , where D(pn , R)′ is the p-power torsion subgroup of D(pn , R). Also ε has order u · pm , where u, which is coprime to p, is the order of the image of ε in D(p, R) and pm is the order of the image in D(pn , R)′ . If m = 0, then the order of ε is coprime to p, so ε is in D(p, R), which means that n = 1, so n = m + 1, as required. If m > 0, then ζ ∈ R must have order divisible by p, so R has characteristic not equal to p. The conductor of ε does not change if we adjoin roots of unity to R, so in light of Lemma 4.2 we may assume that D(N, R) ≈ (Z/N Z)∗ . It follows ′ ′ that for each n′ ≤ n, the p-power subgroup D(pn , R)′ of D(pn , R) is the ′ pn −1 -torsion subgroup of D(pn , R)′ . Thus m = n − 1, since D(pn , R)′ is by assumption the smallest such group that contains the projection of ε. This proves the formula of step (3). We leave the argument when p = 2 as an exercise (see Exercise 4.5). ¤ Example 4.20. If ε = [1, −1, ζ 5 ] ∈ D(200, C), then as in Example 4.17, ε is the product of ε1 = [1, −1] and ε2 = [ζ 5 ]. Because ε1 (5) = −1, the conductor of ε1 is 8. The order of ε2 is 4 (since ζ is a 20th root of unity), so the conductor of ε2 is 5. Thus the conductor of ε is 40 = 8 · 5.

4.5. The Kronecker Symbol In this section all characters have values in C. Frequently ¡ ¢ quadratic characters are described in terms of the Kronecker symbol na , which we define for any integer a and positive integer n as

73

4.5. The Kronecker Symbol

follows. First, if n = p is an odd prime, then for any integer a,  µ ¶   0 if gcd(a, p) 6= 1, a = 1 if a is a square mod p,  p  −1 if a is not a square mod p.

If p = 2, then

   0 if a is even, = 1 if a ≡ ±1 (mod 8),  2  −1 if a ≡ ±3 (mod 8). Q ei More generally, if n = pi with the pi prime, then ³ a ´ Y µ a ¶ei . = n pi ¡a¢ Remark 4.21. One can also ³extend n to n < 0, but we will not need ´ ¡ ¢ a this. The extension is to set −1 = −1 and a1 = 1, for a 6= 0, and to ³ ´ • is extend multiplicatively (in the denominator). Note that the map −1 not a Dirichlet character (see Exercise 4.1). ³a´

Let M be the product of the primes p such that ordp (n) is odd. If M is odd, let N = M ; otherwise, let N = 8M . Lemma 4.22. The function ε(a) =

(¡ ¢ a n

0

if gcd(a, N ) = 1, otherwise

is a Dirichlet character of modulus N . The function (¡ ¢ −1 if a is odd, a ε(a) = 0 if a is even is a Dirichlet character of modulus N . ³ ´ Proof. When restricted to (Z/N Z)∗ , each map p• , for p prime, is a homomorphism, so ε a product of homomorphisms. The second statement follows from the definition and the fact that −1 is a square modulo an odd prime p if and only if p ≡ 1 (mod 4). ¤ This section is about going between representing quadratic characters as row matrices and via Kronecker symbols. This is valuable because the algorithms in [Coh93, §1.1.4] for computing Kronecker symbols run in time

74

4. Dirichlet Characters

quadratic in the number of digits of the input. They³ do ´ not require computa ing discrete logarithms; instead, they use, e.g., that p ≡ a(p−1)/2 (mod p), when p is an odd prime. Algorithm 4.23 (Kronecker Symbol as Dirichlet Character). Given ¡ n¢ > 0, this algorithm computes a representation of the Kronecker symbol n• as a Dirichlet character. (1) [Modulus] Compute N as in Lemma 4.22. (2) [Minimal Generators] Compute minimal generators gi of (Z/N Z)∗ using Algorithm 4.4. ¡ gi ¢ for each gi using one of the algorithms of (3) [Images] Compute N [Coh93, §1.1.4]. ¡ • ¢ Example 4.24. We compute ¡the ¢Dirichlet character associated to 200 . gi Using SAGE, we compute the 200 , for i = 0, 1, 2, where the gi are as in Example 4.9: sage: kronecker(151,200) 1 sage: kronecker(101,200) -1 sage: kronecker(177,200) 1 Thus the corresponding character is defined by [1, −1, 1]. Example 4.25. We compute the character associated to 420 = 4 · 3 · 5 · 7, and minimal generators are g0 = 211,

g1 = 1,

g2 = 281,

g3 = 337,

¡

• 420

¢

. We have

g4 = 241.

We have g0 ≡ −1¡(mod 4), g3 ≡ ¡ gg12¢ ≡ 2 (mod¡ 3), ¡ g23 ¢(mod ¡ g5) ¢ and g4 ≡ 3 g0 ¢ g2 ¢ 4 (mod 7). We find 420 = 420 = 1 and 420 = 420 = 420 = −1. The corresponding character is [1, 1, −1, −1, −1]. Using the following algorithm, we can go in the other direction, i.e., write any quadratic Dirichlet character as a Kronecker symbol. Algorithm 4.26 (Dirichlet Character as Kronecker Symbol). Given ε of order 2 with modulus N , this algorithm writes ε as a Kronecker symbol. (1) [Conductor] Use Algorithm ´ to compute the conductor f of ε. ³ 4.19 • (2) [Odd] If f is odd, output f . ³ ´ ³ ´ ¡ ¢ (3) [Even] If ε(−1) = 1, output f• ; if ε(−1) = −1, output f• · −1 • .

4.6. Restriction, Extension, and Galois Orbits

75

Proof. Since f is the conductor of a quadratic Dirichlet character, it is a square-free product g of odd primes times either 4 or 8, so the group (Z/f Z)∗ does not inject into (Z/gZ)∗ for any proper divisor g of f (see this by reducing to the prime power case). Since g is odd and square-free, ³ ´ • the character g has conductor g. For each odd prime p, by step (3) of ³ ´ Algorithm 4.19 the factor at p of both ε and g• is a quadratic character with modulus p. By Exercise 4.2 and Lemma 4.2 the group D(p,³C) ´ is cyclic, • so it has a unique element of order 2, so the factors of ε and g at p are equal. The quadratic characters with conductor a power of 2 are [−1], ¡ −1[1, ¢ −1], ¡•¢ and [−1, −1]. The character [1, −1] is 2 and the character [−1] is • . ¤ Example 4.27. Consider ε = [−1, −1, −1, −1, −1] with modulus 840 = 8·3· 5·7. It has conductor ¡ a ¢ 840, ¡ −1 and ¢ ε(−1) = −1, so for all a with gcd(a, 840) = 1, we have ε(a) = 840 · a .

4.6. Restriction, Extension, and Galois Orbits

The following two algorithms restrict and extend characters to a compatible modulus. Using them, it is easy to define multiplication of two characters ε ∈ D(N, R) and ε′ ∈ D(N ′ , R′ ), as long as R and R′ are subrings of a common ring. To carry out the multiplication, extend both characters to a common base ring, and then extend them to characters modulo lcm(N, N ′ ) and multiply. Algorithm 4.28 (Restriction of Character). Given a Dirichlet character ε ∈ D(N, R) and a divisor N ′ of N that is a multiple of the conductor of ε, this algorithm finds a characters ε′ ∈ D(N ′ , R), such that ε′ (a) = ε(a), for all a ∈ Z with (a, N ) = 1. (1) [Conductor] Compute the conductor of ε using Algorithm 4.19, and verify that N ′ is divisible by the conductor and divides N . (2) [Minimal Generators] Compute minimal generators gi for (Z/N ′ Z)∗ . (3) [Values of Restriction] For each i, compute ε′ (gi ) as follows. Find a multiple aN ′ of N ′ such that (gi + aN ′ , N ) = 1; then ε′ (gi ) = ε(gi + aN ′ ). (4) [Output Character] Output the Dirichlet character of modulus N ′ defined by [ε′ (g0 ), . . . , ε′ (gn )]. Proof. The only part that is not clear is that in step (3) there is an a such that (gi + aN ′ , N ) = 1. If we write N = N1 · N2 , with (N1 , N2 ) = 1 and N1 divisible by all primes that divide N ′ , then (gi , N1 ) = 1 since (gi , N ′ ) = 1. By the Chinese Remainder Theorem, there is an x ∈ Z such that x ≡ gi

76

4. Dirichlet Characters

(mod N1 ) and x ≡ 1 (mod N2 ). Then x = gi + bN1 = gi + (bN1 /N ′ ) · N ′ and (x, N ) = 1, which completes the proof. ¤ Algorithm 4.29 (Extension of Character). Given a Dirichlet character ε ∈ D(N, R) and a multiple N ′ of N , this algorithm finds a character ε′ ∈ D(N ′ , R), such that ε′ (a) = ε(a), for all a ∈ Z with (a, N ′ ) = 1.

(1) [Minimal Generators] Compute minimal generators gi for (Z/N ′ Z)∗ . (2) [Evaluate] Compute ε(gi ) for each i. Since (gi , N ′ ) = 1, we also have (gi , N ) = 1. (3) [Output Character] Output the character [ε(g0 ), . . . , ε(gn )].

Let F be the prime subfield of R, and assume that R ⊂ F , where F is a separable closure of F . If σ ∈ Gal(F /F ) and ε ∈ D(N, R), let (σε)(n) = σ(ε(n)); this defines an action of Gal(F /F ) on D(N, R). Our next algorithm computes the orbits for the action of Gal(F /F ) on D(N, R). This algorithm can provide huge savings for modular forms computations because the spaces Mk (N, ε) and Mk (N, ε′ ) are canonically isomorphic if ε and ε′ are conjugate. Algorithm 4.30 (Galois Orbit). Given a Dirichlet character ε ∈ D(N, R), this algorithm computes the orbit of ε under the action of G = Gal(F /F ), where F is the prime subfield of Frac(R), so F = Fp or Q. (1) [Order of ζ] Let n be the order of the chosen root ζ ∈ R. (2) [Nontrivial Automorphisms] If char(R) = 0, let A = {a : 2 ≤ a < n and (a, n) = 1}. If char(R) = p > 0, compute the multiplicative order r of p (mod n), and let A = {pm : 1 ≤ m < r}.

(3) [Compute Orbit] Compute and output the set of unique elements εa for each a ∈ A (there could be repeats, so we output unique elements only). Proof. We prove that the nontrivial automorphisms of hζi in characteristic p are as in step (2). It is well known that every automorphism in s characteristic p on ζ ∈ Fp is of the form x 7→ xp , for some s. The images of ζ under such automorphisms are 2

ζ, ζ p , ζ p , . . . . r

Suppose r > 0 is minimal such that ζ = ζ p . Then the orbit of ζ is r−1 ζ, . . . , ζ p . Also pr ≡ 1 (mod n), where n is the multiplicative order of ζ, so r is the multiplicative order of p modulo n, which completes the proof. ¤

4.7. Alternative Representations of Characters

77

Example 4.31. The Galois orbits of characters in D(20, C∗ ) are as follows: G0 = {[1, 1, 1]},

G1 = {[−1, 1, 1]},

G2 = {[1, 1, ζ4 ], [1, 1, −ζ4 ]}

G3 = {[−1, 1, ζ4 ], [−1, 1, −ζ4 ]}

G4 = {[1, 1, −1]},

G5 = {[−1, 1, −1]}.

The conductors of the characters in orbit G0 are 1, in orbit G1 they are 4, in orbit G2 they are 5, in G3 they are 20, in G4 the conductor is 5, and in G5 the conductor is 20. (You should verify this.) SAGE computes Galois orbits as follows: sage: G = DirichletGroup(20) sage: G.galois_orbits() [ [[1, 1]], [[1, zeta4], [1, -zeta4]], [[1, -1]], [[-1, 1]], [[-1, zeta4], [-1, -zeta4]], [[-1, -1]] ]

4.7. Alternative Representations of Characters Let N be a positive integer and R an integral domain, with fixed root of unity ζ of order n, Q and let D(N, R) = D(N, R, ζ). As in the rest of this chapter, write N = pei i , and let Ci = hgi i be the corresponding cyclic factors of (Z/N Z)∗ . In this section we discuss other ways to represent elements ε ∈ D(N, R). Each representation has advantages and disadvantages, and no single representation is best. It is easy to convert between them, and some algorithms are much easier using one representation than when using another. In this section we present two other representations, each having advantages and disadvantages. There is no reason to restrict to only one representation; for example, SAGE internally uses both. We could represent ε by giving a list [b0 , . . . , br ], where each bi ∈ Z/nZ and ε(gi ) = ζ bi . Then arithmetic in D(N, R) is arithmetic in (Z/nZ)r+1 , which is very efficient. A drawback to this approach (in practice) is that it is easy to accidentally consider sequences that do not actually correspond to

78

4. Dirichlet Characters

elements of D(N, R). Also the choice of ζ is less clear, which can cause confusion. Finally, the orders of the local factors is more opaque, e.g., compare [−1, ζ40 ] with [20, 1]. Overall this representation is not too bad and is more like representing a linear transformation by a matrix. It has the advantage over the representation discussed earlier in this chapter that arithmetic in D(N, R) is very efficient and does not require any operations in the ring R. Another way to represent ε would be to give a list [b0 , . . . , br ] of integers, but this time with bi ∈ Z/ gcd(si , n)Z, where si is the order of gi . Then ε(gi ) = ζ bi ·n/(gcd(si ,n)) ,

which is already pretty complicated. With this representation we set up an identification M D(N, R) ∼ Z/ gcd(si , n)Z, = i

and arithmetic is efficient. This approach is seductive because every sequence of integers determines a character, and the sizes of the integers in the sequence nicely indicate the local orders of the character. However, giving analogues of many of the algorithms discussed in this chapter that operate on characters represented this way is tricky. For example, the representation depends very much on the order of ζ, so it is difficult to correctly compute natural maps D(N, R) → D(N, S), for R ⊂ S rings.

4.8. Dirichlet Characters in SAGE To create a Dirichlet character in SAGE, first create the group D(N, R) of Dirichlet characters then construct elements of that group. First we make D(11, Q): sage: G = DirichletGroup(11, QQ); G Group of Dirichlet characters of modulus 11 over Rational Field A Dirichlet character prints as a matrix that gives the values of the character on canonical generators of (Z/N Z)∗ (as discussed below). sage: [[1], sage: sage: [-1]

list(G) [-1]] eps = G.0 eps

# 0th generator for Dirichlet group

The character ε takes the value −1 on the unit generator.

4.8. Dirichlet Characters in SAGE

79

sage: G.unit_gens() [2] sage: eps(2) -1 sage: eps(3) 1 It is 0 on any integer not coprime to 11: sage: [eps(11*n) for n in range(10)] [0, 0, 0, 0, 0, 0, 0, 0, 0, 0] We can also create groups of Dirichlet characters taking values in other rings or fields. For example, we create the cyclotomic field Q(ζ4 ). sage: R = CyclotomicField(4) sage: CyclotomicField(4) Cyclotomic Field of order 4 and degree 2 Then we define G = D(15, Q(ζ4 )). sage: G = DirichletGroup(15, R) sage: G Group of Dirichlet characters of modulus 15 over Cyclotomic Field of order 4 and degree 2 Next we list each of its elements. sage: list(G) [[1, 1], [-1, 1], [1, zeta4], [-1, zeta4], [1, -1], [-1, -1], [1, -zeta4], [-1, -zeta4]] Now we evaluate the second generator of G on various integers:

80

4. Dirichlet Characters

sage: sage: -1 sage: -1 sage: 0

e = G.1 e(4) e(-1) e(5)

Finally we list all the values of e. sage: [e(n) for n in range(15)] [0, 1, zeta4, 0, -1, 0, 0, zeta4, -zeta4, 0, 0, 1, 0, -zeta4, -1] We can also compute with groups of Dirichlet characters with values in a finite field. sage: G = DirichletGroup(15, GF(5)); G Group of Dirichlet characters of modulus 15 over Finite Field of size 5 We list all the elements of G, again represented by lists that give the images of each unit generator, as an element of F5 . sage: list(G) [[1, 1], [4, 1], [1, 2], [4, 2], [1, 4], [4, 4], [1, 3], [4, 3]] We evaluate the second generator of G on several integers. sage: e = G.1 sage: e(-1) 4 sage: e(2) 2 sage: e(5) 0 sage: print [e(n) for n in range(15)] [0, 1, 2, 0, 4, 0, 0, 2, 3, 0, 0, 1, 0, 3, 4]

4.9. Exercises

81

4.9. Exercises 4.1 Let f : Z → C be the map given by    0 if a = 0, f (a) = −1 if a < 0,   1 if a > 0.

Prove that f is not a Dirichlet character of any modulus N .

4.2 This exercise is about the structure of the units of Z/pn Z. (a) If p is odd and n is a positive integer, prove that (Z/pn Z)∗ is cyclic. (b) For n ≥ 3, prove that (Z/2n Z)∗ is a direct sum of the cyclic subgroups h−1i and h5i, of orders 2 and 2n−2 , respectively. 4.3 Prove that Algorithm 4.4 works, i.e., that if g ∈ (Z/pr Z)∗ and g n/pi 6= 1 for all pi | n = ϕ(pr ), then g is a generator of (Z/pr Z)∗ . 4.4 (a) Let p be an odd prime and n ≥ 2 an integer, and prove that ¡ ¢ (1 + pn−1 Z/pn Z), × ∼ = (Z/pZ, +).

(b) Use the first part to show that solving the discrete log problem in (Z/pn Z)∗ is “not much harder” than solving the discrete log problem in (Z/pZ)∗ .

4.5 Suppose ε is a nontrivial Dirichlet character of modulus 2n of order r over the complex numbers C. Prove that the conductor of ε is ( 2ord2 (r)+1 if ε(5) = 1, c= 2ord2 (r)+2 if ε(5) 6= 1.

4.6 (a) Find an irreducible quadratic polynomial f over F5 . (b) Then F25 = F5 [x]/(f ). Find an element with multiplicative order 4 in F25 . (c) Make a list of all Dirichlet characters in D(25, F25 , ζ). (d) Divide these characters into orbits for the action of Gal(F5 /F5 ).

Chapter 5

Eisenstein Series and Bernoulli Numbers

We introduce generalized Bernoulli numbers attached to Dirichlet characters and give an algorithm to enumerate the Eisenstein series in Mk (N, ε).

5.1. The Eisenstein Subspace Let Mk (Γ1 (N )) be the space of modular forms of weight k for Γ1 (N ), and let T be the Hecke algebra acting on Mk (Γ1 (N )), which is the subring of End(Mk (Γ1 (N ))) generated by all Hecke operators. Then there is a Tmodule decomposition Mk (Γ1 (N )) = Ek (Γ1 (N )) ⊕ Sk (Γ1 (N )), where Sk (Γ1 (N )) is the subspace of modular forms that vanish at all cusps and Ek (Γ1 (N )) is the Eisenstein subspace, which is uniquely determined by this decomposition. The above decomposition induces a decomposition of Mk (Γ0 (N )) and of Mk (N, ε), for any Dirichlet character ε of modulus N .

5.2. Generalized Bernoulli Numbers Suppose ε is a Dirichlet character of modulus N over C. Leopoldt [Leo58] defined generalized Bernoulli numbers attached to ε. Definition 5.1 (Generalized Bernoulli Number). We define the generalized Bernoulli numbers Bk,ε attached to ε by the following identity of infinite 83

84

5. Eisenstein Series and Bernoulli Numbers

series: N X ε(a) · x · eax a=1

eN x − 1

=

∞ X k=0

Bk,ε ·

xk . k!

If ε is the trivial character of modulus 1 and Bk are as in Section 2.1, then Bk,ε = Bk , except when k = 1, in which case B1,ε = −B1 = 1/2 (see Exercise 5.2).

5.2.1. Algebraically Computing Generalized Bernoulli Numbers. Let Q(ε) denote the field generated by the image of the character ε; thus Q(ε) is the cyclotomic extension Q(ζn ), where n is the order of ε. Algorithm 5.2 (Generalized Bernoulli Numbers). Given an integer k ≥ 0 and any Dirichlet character ε with modulus N , this algorithm computes the generalized Bernoulli numbers Bj,ε , for j ≤ k. N x −1) ∈ Q[[x]] to precision O(xk+1 ) by comput(1) Compute g = x/(e P N x ing e − 1 = n≥1 N n xn /n! to precision O(xk+2 ) and computing the inverse 1/(eN x − 1), then multiplying by x. (2) For each a = 1, . . . , N , compute fa = g · eax ∈PQ[[x]], to precin n sion O(xk+1 ). This requires computing eax = n≥0 a x /n! to k+1 N x precision O(x ). (Omit computation of e if N > 1 since then ε(N ) = 0.) (3) Then for j ≤ k, we have

Bj,ε = j! ·

N X a=1

ε(a) · cj (fa ),

where cj (fa ) is the coefficient of xj in fa . Note that in steps (1) and (2) we compute the power series doing arithmetic only in Q[[x]], not in Q(ε)[[x]], which could be much less efficient if ε has large order. In step (1) if k is huge, we could compute the inverse 1/(eN x − 1) using asymptotically fast arithmetic and Newton iteration. Example 5.3. The nontrivial character ε with modulus 4 has order 2 and takes values in Q. The Bernoulli numbers Bk,ε for k even are all 0 and for

5.2. Generalized Bernoulli Numbers

85

k odd they are B1,ε = −1/2,

B3,ε = 3/2,

B5,ε = −25/2,

B7,ε = 427/2,

B9,ε = −12465/2,

B11,ε = 555731/2,

B13,ε = −35135945/2,

B15,ε = 2990414715/2, B17,ε = −329655706465/2,

B19,ε = 45692713833379/2. Example 5.4. The generalized Bernoulli numbers√need not be in Q. Suppose ε is the mod 5 character such that ε(2) = i = −1. Then Bk,ε = 0 for k even and B1,ε = B3,ε = B5,ε = B7,ε = B9,ε = B11,ε = B13,ε = B15,ε = B17,ε =

−i − 3 , 5 6i + 12 , 5 −86i − 148 , 5 2366i + 3892 , 5 −108846i − 176868 , 5 7599526i + 12309572 , 5 −751182406i − 1215768788 , 5 99909993486i + 161668772052 , 5 −17209733596766i − 27846408467908 . 5

Example 5.5. We use SAGE to compute some of the above generalized Bernoulli numbers. First we √ define the character and verify that ε(2) = i (note that in SAGE zeta4 is −1).

86

5. Eisenstein Series and Bernoulli Numbers

sage: G = DirichletGroup(5) sage: e = G.0 sage: e(2) zeta4 We compute the Bernoulli number B1,ε . sage: e.bernoulli(1) -1/5*zeta4 - 3/5 We compute B9,ε . sage: e.bernoulli(9) -108846/5*zeta4 - 176868/5 Proposition 5.6. If ε(−1) 6= (−1)k and k ≥ 2, then Bk,ε = 0. Proof. See Exercise 5.3.

¤

5.2.2. Computing Generalized Bernoulli Numbers Analytically. This section, which was written jointly with Kevin McGown, is about a way to compute generalized Bernoulli numbers, which is similar to the algorithm in Section 2.7. Let χ be a primitive Dirichlet character modulo its conductor f . Note from the definition of Bernoulli numbers that if σ ∈ Gal(Q/Q), then (5.2.1)

σ(Bn,χ ) = Bn,σ(χ) .

For any character χ, we define the Gauss sum τ (χ) as τ (χ) =

f −1 X

χ(r) ζ r ,

r=1

where ζ = exp(2πi/f ) is the principal f th root of unity. The Dirichlet L-function for χ for Re(s) > 1 is ∞ X χ(n) n−s . L(s, χ) = n=1

In the right half plane {s ∈ C | Re(s) > 1} this function is analytic, and because χ is multiplicative, we have the Euler product representation Y ¡ ¢−1 (5.2.2) . L(s, χ) = 1 − χ(p)p−s p prime

We note (but will not use) that through analytic continuation L(s, χ) can be extended to a meromorphic function on the entire complex plane.

87

5.2. Generalized Bernoulli Numbers

If χ is a nonprincipal primitive Dirichlet character of conductor f such that χ(−1) = (−1)n , then (see, e.g., [Wan82]) µ ¶ 2πi n Bn,χ n−1 τ (χ) . L(n, χ) = (−1) 2 f n! Solving for the Bernoulli number yields n−1

Bn,χ = (−1)

2n! τ (χ)

µ

f 2πi

¶n

L(n, χ) .

This allows us to give decimal approximations for Bn,χ . It remains to compute Bn,χ exactly (i.e., as an algebraic integer). To simplify the above expression, we define µ ¶n f n−1 Kn,χ = (−1) 2n! 2i and write (5.2.3)

Bn,χ =

Kn,χ L(n, χ) . π n τ (χ)

Note that we can compute Kn,χ exactly in the field Q(i). The following result identifies the denominator of Bn,χ . Theorem 5.7. Let n and χ   1      2 d= 1   np     (1 − χ(1 + p))

be as above, and define an integer d as follows: if if if if if

f f f f f

is divisible by two distinct primes, = 4, = 2µ , µ > 2, = p, p > 2, = pµ , p > 2, µ > 1.

Then dn−1 Bn,χ is integral.

Proof. See [Car59a] for the proof and [Car59b] for further details.

¤

To compute the algebraic integer dn−1 Bn,χ , and we compute L(n, χ) to very high precision using the Euler product (5.2.2) and the formula (5.2.3). We carry out the same computation for each of the Gal(Q/Q) conjugates of χ, which by (5.2.1) yields the conjugates of dn−1 Bn,χ . We can then write down the characteristic polynomial of dn−1 Bn,χ to very high precision and recognize the coefficients as rational integers. Finally, we determine which of the roots of the characteristic polynomial is dn−1 Bn,χ by approximating them all numerically to high precision and seeing which is closest to our numerical approximation to dn−1 Bn,χ . The details are similar to what is explained in Section 2.7.

88

5. Eisenstein Series and Bernoulli Numbers

5.3. Explicit Basis for the Eisenstein Subspace Suppose χ and ψ are primitive Dirichlet characters with conductors L and R, respectively. Let   X X  (5.3.1) Ek,χ,ψ (q) = c0 + ψ(n) · χ(m/n) · nk−1  q m ∈ Q(χ, ψ)[[q]], m≥1

where

n|m

 0

if L > 1, Bk,ψ − if L = 1. 2k Note that when χ = ψ = 1 and k ≥ 4, then Ek,χ,ψ = Ek , where Ek is from Chapter 1. c0 =

Miyake proves statements that imply the following in [Miy89, Ch. 7]. Theorem 5.8. Suppose t is a positive integer and χ, ψ are as above and that k is a positive integer such that χ(−1)ψ(−1) = (−1)k . Except when k = 2 and χ = ψ = 1, the power series Ek,χ,ψ (q t ) defines an element of Mk (RLt, χψ). If χ = ψ = 1, k = 2, t > 1, and E2 (q) = Ek,χ,ψ (q), then E2 (q) − tE2 (q t ) is a modular form in M2 (Γ0 (t)). Theorem 5.9. The Eisenstein series in Mk (N, ε) coming from Theorem 5.8 with RLt | N and χψ = ε form a basis for the Eisenstein subspace Ek (N, ε). Theorem 5.10. The Eisenstein series Ek,χ,ψ (q) ∈ Mk (RL) defined above are eigenforms (i.e., eigenvectors for all Hecke operators Tn ). Also E2 (q) − tE2 (q t ), for t > 1, is an eigenform. Since Ek,χ,ψ (q) is normalized so the coefficient of q is 1, the eigenvalue of Tm is the coefficient X ψ(n) · χ(m/n) · nk−1 n|m

qm

of (see Proposition 9.10). Also for f = E2 (q) − tE2 (q t ) with t > 1 prime, the coefficient of q is 1, Tm (f ) = σ1 (m) · f for (m, t) = 1, and Tt (f ) = ((t + 1) − t)f = f . Algorithm 5.11 (Enumerating Eisenstein Series). Given a weight k and a Dirichlet character ε of modulus N , this algorithm computes a basis for the Eisenstein subspace Ek (N, ε) of Mk (N, ε) to precision O(q r ). (1) [Weight 2 Trivial Character?] If k = 2 and ε = 1, output the Eisenstein series E2 (q) − tE2 (q t ), for each divisor t | N with t = 6 1, and then terminate.

5.3. Explicit Basis for the Eisenstein Subspace

89

(2) [Empty Space?] If ε(−1) 6= (−1)k , output the empty list. (3) [Compute Dirichlet Group] Let G = D(N, Q(ζn )) be the group of Dirichlet characters with values in Q(ζn ), where n is the exponent of (Z/N Z)∗ . (4) [Compute Conductors] Compute the conductor of every element of G using Algorithm 4.19. (5) [List Characters χ] Form a list V of all Dirichlet characters χ ∈ G such that cond(χ) · cond(χ/ε) divides N . (6) [Compute Eisenstein Series] For each character χ in V , let ψ = χ/ε and compute Ek,χ,ψ (q t ) (mod q r ) for each divisor t of N/(cond(χ) · cond(ψ)). Here we compute Ek,χ,ψ (q t ) (mod q r ) using (5.3.1) and Algorithm 5.2. Remark 5.12. Algorithm 5.11 is what is currently used in SAGE. It might be better to first reduce to the prime power case by writing all characters as a product of local characters and combine steps (4) and (5) into a single step that involves orders. However, this might make things more obscure. Example 5.13. The following is a basis of Eisenstein series for E2 (Γ1 (13)). 1 + q + 3q 2 + 4q 3 + · · · , 2 ¡ 2 ¢ ¡ ¢ 11 7 2 2 − + q + 2ζ12 + 1 q 2 + −3ζ12 + 1 q3 + · · · , = − ζ12 13¡ 13¢ ¡ 2 ¢ 2 = q + ζ12 + 2 q 2 + −ζ12 + 3 q3 + · · · , ¡ 2 ¢ ¡ 2 ¢ 2 = −ζ12 + q + 2ζ12 − 1 q 2 + 3ζ12 − 2 q3 + · · · , ¡ 2 ¢ ¡ 2 ¢ = q + ζ12 + 1 q 2 + ζ12 + 2 q3 + · · · ,

f1 = f2 f3 f4 f5

f6 = −1 + q + −q 2 + 4q 3 + · · · ,

f7 = q + q 2 + 4q 3 + · · · , ¡ ¢ ¡ ¢ 2 2 2 f8 = ζ12 − 1 + q + −2ζ12 + 1 q 2 + −3ζ12 + 1 q3 + · · · , ¡ 2 ¢ ¡ 2 ¢ f9 = q + −ζ12 + 2 q 2 + −ζ12 + 3 q3 + · · · , ¡ ¢ ¡ 2 ¢ 7 2 18 2 f10 = ζ12 − + q + −2ζ12 + 3 q 2 + 3ζ12 − 2 q3 + · · · , 13 ¡ 13 ¢ ¡ 2 ¢ 2 f11 = q + −ζ12 + 3 q 2 + ζ12 + 2 q3 + · · · . We computed it as follows: sage: E = EisensteinForms(Gamma1(13),2) sage: E.eisenstein_series()

90

5. Eisenstein Series and Bernoulli Numbers

We can also compute the parameters χ, ψ, t that define each series: sage: e = E.eisenstein_series() sage: for e in E.eisenstein_series(): ... print e.parameters() ... ([1], [1], 13) ([1], [zeta6], 1) ([zeta6], [1], 1) ([1], [zeta6 - 1], 1) ([zeta6 - 1], [1], 1) ([1], [-1], 1) ([-1], [1], 1) ([1], [-zeta6], 1) ([-zeta6], [1], 1) ([1], [-zeta6 + 1], 1) ([-zeta6 + 1], [1], 1)

5.4. Exercises 5.1 Suppose A and B are diagonalizable linear transformations of a finite-dimensional vector space V over an algebraically closed field K and that AB = BA. Prove there is a basis for V so that the matrices of A and B with respect to that basis are both simultaneously diagonal. 5.2 If ε is the trivial character of modulus 1 and Bk are as in Section 2.1, then Bk,ε = Bk , except when k = 1, in which case B1,ε = −B1 = 1/2. 5.3 Prove that for k ≥ 2 if ε(−1) 6= (−1)k , then Bk,ε = 0.

5.4 Show that the dimension of the Eisenstein subspace E3 (Γ1 (13)) is 12 by finding a basis of series Ek,χ,ψ . You do not have to write down the q-expansions of the series, but you do have to figure out which χ, ψ to use.

Chapter 6

Dimension Formulas

When computing with spaces of modular forms, it is helpful to have easy-tocompute formulas for dimensions of these spaces. Such formulas provide a check on the output of the algorithms from Chapter 8 that compute explicit bases for spaces of modular forms. We can also use dimension formulas to improve the efficiency of some of the algorithms in Chapter 8, since we can use them to determine the ranks of certain matrices without having to explicitly compute those matrices. Dimension formulas can also be used in generating bases of q-expansions; if we know the dimension of Mk (N, ε) and if we have a process for computing q-expansions of elements of Mk (N, ε), e.g., multiplying together q-expansions of certain forms of smaller weight, then we can tell when we are done generating Mk (N, ε). This chapter contains formulas for dimensions of spaces of modular forms, along with some remarks about how to evaluate these formulas. In some cases we give dimension formulas for spaces that we will define in later chapters. We also give many examples, some of which were computed using the modular symbols algorithms from Chapter 8. Many of the dimension formulas and algorithms we give below grew out of Shimura’s book [Shi94] and a program that Bruce Kaskel wrote (around 1996) in PARI, which Kevin Buzzard extended. That program codified dimension formulas that Buzzard and Kaskel found or extracted from the literature (mainly [Shi94, §2.6]). The algorithms for dimensions of spaces with nontrivial character are from [CO77], with some refinements suggested by Kevin Buzzard. For the rest of this chapter, N denotes a positive integer and k ≥ 2 is an integer. We will give no simple formulas for dimensions of spaces of weight 1 modular forms; in fact, it might not be possible to give such formulas since 91

92

6. Dimension Formulas

the methods used to derive the formulas below do not apply in the case k = 1. If k = 0, the only modular forms are the constants, and for k < 0 the dimension of Mk (N, ε) is 0. For a nonzero integer N and a prime p, let vp (N ) be the largest integer e such that pe | N . In the formulas in this chapter, p always denotes a prime number. Let Mk (N, ε) be the space of modular forms of level N weight k and character ε, and let Sk (N, ε) and Ek (N, ε) be the cuspidal and Eisenstein subspaces, respectively. The dimension formulas below for Sk (Γ0 (N )), Sk (Γ1 (N )), Ek (Γ0 (N )) and Ek (Γ1 (N )) can be found in [DS05, Ch. 3], [Shi94, §2.6]1 and [Miy89, §2.5]. They are derived using the Riemann-Roch Theorem applied to the covering X0 (N ) → X0 (1) or X1 (N ) → X1 (1) and appropriately chosen divisors. It would be natural to give a sample argument along these lines at this point, but we will not since it easy to find such arguments in other books and survey papers (see, e.g., [DI95]). So you will not learn much about how to derive dimension formulas from this chapter. What you will learn is precisely what the dimension formulas are, which is something that is often hard to extract from obscure references. In addition to reading this chapter, the reader may wish to consult [Mar05] for proofs of similar dimension formulas, asymptotic results, and a nonrecursive formula for dimensions of certain new subspaces.

6.1. Modular Forms for Γ0 (N ) For any prime p and any positive integer N , let vp (N ) be the power of p that divides N . Also, let ´ Y³ vp (N ) vp (N )−1 µ0 (N ) = p +p , p|N

( 0 µ0,2 (N ) = Q

( 0 µ0,3 (N ) = Q c0 (N ) =

X

p|N

p|N

³

³

1+

1+

³

³

−4 p

−3 p

´´ ´´

if 4 | N ,

otherwise, if 2 | N or 9 | N ,

otherwise,

ϕ(gcd(d, N/d)),

d|N

g0 (N ) = 1 +

µ0,3 (N ) c0 (N ) µ0 (N ) µ0,2 (N ) − − − . 12 4 3 2

1The formulas in [Shi94, §2.6] contain some minor mistakes.

6.1. Modular Forms for Γ0 (N )

93

Note that µ0 (N ) is the index of Γ0 (N ) in SL2 (Z) (see Exercise 6.1). Proposition 6.1. We have dim S2 (Γ0 (N )) = g0 (N ), and for k ≥ 4 even, µ ¶ k dim Sk (Γ0 (N )) = (k − 1) · (g0 (N ) − 1) + − 1 · c0 (N ) 2 ¹ º ¹ º k k + µ0,3 (N ) · . + µ0,2 (N ) · 4 3 The dimension of the Eisenstein subspace is ( c0 (N ) if k 6= 2, dim Ek (Γ0 (N )) = c0 (N ) − 1 if k = 2. The following is a table of dim Sk (Γ0 (N )) for some values of N and k: N 1 10 11 100 389 1000 2007 100000

S2 (Γ0 (N )) 0 0 1 7 32 131 221 14801

S4 (Γ0 (N )) 0 3 2 36 97 430 806 44800

S6 (Γ0 (N )) 0 5 4 66 161 730 1346 74800

S24 (Γ0 (N )) 2 33 22 336 747 3430 6206 344800

Example 6.2. Use the commands dimension cusp forms, dimension eis, and dimension modular forms to compute the dimensions of the three spaces Sk (Γ0 (N )), Ek (Γ0 (N )) and Mk (Γ0 (N )), respectively. For example, sage: dimension_cusp_forms(Gamma0(2007),2) 221 sage: dimension_eis(Gamma0(2007),2) 7 sage: dimension_modular_forms(Gamma0(2007),2) 228 Remark 6.3. Csirik, Wetherell, and Zieve prove in [CWZ01] that a random positive integer has probability 0 of being a value of g0 (N ) = dim S2 (Γ0 (N )), and they give bounds on the size of the set of values of g0 (N ) below some given x. For example, they show that 150, 180, 210, 286, 304, 312, . . . are the first few integers that are not of the form g0 (N ) for any N . See Figure 6.1.1 for a plot of the very erratic function g0 (N ). In contrast, the function k 7→ dim S2k (Γ0 (12)) is very well behaved (see Figure 6.1.2).

94

6. Dimension Formulas

Figure 6.1.1. Dimension of S2 (Γ0 (N )) as a function of N .

Figure 6.1.2. Dimension of S2k (Γ0 (12)) as a function of k.

6.2. Modular Forms for Γ1 (N )

95

6.1.1. New and Old Subspaces. In this section we assume the reader is either familiar with newforms or has read Section 9.2. For any integer R, let   if p3 | R for some p, 0Y µ(R) = −2 otherwise,   p||R

where the product is over primes that exactly divide R. Note that µ is not the Moebius function, but it has a similar flavor. Proposition 6.4. The dimension of the new subspace is dim Sk (Γ0 (N ))new =

X

M |N

µ(N/M ) · dim Sk (Γ0 (M )),

where the sum is over the positive divisors M of N . As a consequence of Theorem 9.4, we also have dim Sk (Γ0 (N )) =

X

σ0 (N/M ) dim Sk (Γ0 (M ))new ,

M |N

where σ0 (N/M ) is the number of divisors of N/M . Example 6.5. We compute the dimension of the new subspace of Sk (Γ0 (N )) using the SAGE command dimension new cusp forms as follows: sage: 8 sage: 10 sage: 1017 sage: 2460

dimension_new_cusp_forms(Gamma0(11),12) dimension_cusp_forms(Gamma0(11),12) dimension_new_cusp_forms(Gamma0(2007),12) dimension_cusp_forms(Gamma0(2007),12)

6.2. Modular Forms for Γ1 (N ) This section follows Section 6.1 closely, but with suitable modifications with Γ0 (N ) replaced by Γ1 (N ).

96

6. Dimension Formulas

Define functions of a positive integer N by the following formulas:  µ0 (N ) if N = 1, 2, µ1 (N ) = φ(N ) · µ0 (N )  otherwise, 2 ( 0 if N ≥ 4, µ1,2 (N ) = µ0,2 (N ) otherwise, ( 0 if N ≥ 4, µ1,3 (N ) = µ0,3 (N ) otherwise,   c0 (N ) if N = 1, 2,    3 if N = 4, c1 (N ) = X φ(d)φ(N/d)   otherwise,    2 d|N

µ1 (N ) µ1,2 (N ) µ1,3 (N ) c1 (N ) − − − . 12 4 3 2 Note that g1 (N ) is the genus of the modular curve X1 (N ) (associated to Γ1 (N )) and c1 (N ) is the number of cusps of X1 (N ). g1 (N ) = 1 +

Proposition 6.6. We have dim S2 (Γ1 (N )) = g1 (N ). If N ≤ 2, then Γ0 (N ) = Γ1 (N ) so dim Sk (Γ1 (N )) = dim Sk (Γ0 (N )), where dim Sk (Γ0 (N )) is given by the formula of Proposition 6.1. If k ≥ 3, let µ ¶ k a(N, k) = (k − 1)(g1 (N ) − 1) + − 1 · c1 (N ). 2 Then for N ≥ 3,   if N = 4 and 2 ∤ k, a + 1/2 dim Sk (Γ1 (N )) = a + ⌊k/3⌋ if N = 3,   a otherwise. The dimension of the Eisenstein subspace is as follows: ( c1 (N ) if k 6= 2, dim Ek (Γ1 (N )) = c1 (N ) − 1 if k = 2.

The dimension of the new subspace of Mk (Γ1 (N )) is X µ(N/M ) · dim Sk (Γ1 (M )), dim Sk (Γ1 (N ))new = M |N

where µ is as in the statement of Proposition 6.4.

6.2. Modular Forms for Γ1 (N )

97

Remark 6.7. Since Mk = Sk ⊕ Ek , the formulas above for dim Sk and dim Ek also yield a formula for the dimension of Mk .

Figure 6.2.1. Dimension of S2 (Γ1 (N )) as a function of N .

The following table contains the dimension of Sk (Γ1 (N )) for some sample values of N and k: N 1 10 11 100 389 1000 2007 100000

S2 (Γ1 (N )) 0 0 1 231 6112 28921 147409 299792001

S3 (Γ1 (N )) 0 2 5 530 12416 58920 296592 599792000

S4 (Γ1 (N )) 0 5 10 830 18721 88920 445776 899792000

S24 (Γ1 (N )) 2 65 110 6830 144821 688920 3429456 6899792000

Example 6.8. We compute dimensions of spaces of modular forms for Γ1 (N ):

98

6. Dimension Formulas

sage: dimension_cusp_forms(Gamma1(2007),2) 147409 sage: dimension_eis(Gamma1(2007),2) 3551 sage: dimension_modular_forms(Gamma1(2007),2) 150960

6.3. Modular Forms with Character Fix a Dirichlet character ε of modulus N , and let c be the conductor of ε (we do not assume that ε is primitive). Assume that ε 6= 1, since otherwise Mk (N, ε) = Mk (Γ0 (N )) and the formulas of Section 6.1 apply. Also, assume that ε(−1) = (−1)k , since otherwise dim Mk (Γ0 (N )) = 0. In this section we discuss formulas for computing each of Mk (N, ε), Sk (N, ε) and Ek (N, ε). In [CO77], Cohen and Oesterl´e assert (without published proof; see Remark 6.11 below) that for any k ∈ Z and N , ε as above, dim Sk (N, ε) − dim M2−k (N, ε) =

k−1 1 Y λ(p, N, vp (c)) · µ0 (N ) − · 12 2 p|N X X ε(x) + γ3 (k) · + γ4 (k) ·

ε(x)

x∈A3 (N )

x∈A4 (N )

where µ0 (N ) is as in Section 6.1, A4 (N ) = {x ∈ Z/N Z : x2 + 1 = 0} and A3 (N ) = {x ∈ Z/N Z : x2 + x + 1 = 0}, and γ3 , γ4 are   −1/4 if k ≡ 2 (mod γ4 (k) = 1/4 if k ≡ 0 (mod   0 if k is odd.   −1/3 if k ≡ 2 (mod γ3 (k) = 1/3 if k ≡ 0 (mod   0 if k ≡ 1 (mod

4), 4), 3), 3), 3).

It remains to define λ. Fix a prime divisor p | N and let r = vp (N ). Then  r r −1  p 2 + p 2 r−1 λ(p, N, vp (c)) = 2 · p 2   2 · pr−vp (c)

if 2 · vp (c) ≤ r and 2 | r, if 2 · vp (c) ≤ r and 2 ∤ r, if 2 · vp (c) > r.

99

6.3. Modular Forms with Character

This flexible formula can be used to compute the dimension of Mk (N, ε), Sk (N, ε), and Ek (N, ε) for any N , ε, k 6= 1, by using that dim Sk (N, ε) = 0 dim Mk (N, ε) = 0 dim M0 (N, ε) = 1

if k ≤ 0,

if k < 0,

if k = 0.

One thing that is not straightforward when implementing an algorithm to compute the above dimension formulas is how to efficiently compute the sets A4 (N ) and A6 (N ). Kevin Buzzard suggested the following two algorithms. Note that if k is odd, then γ4 (k) = 0, so the sum over A4 (N ) is only needed when k is even. Algorithm 6.9 (Sum over A4 (N )). Given a positive integer P N and an even Dirichlet character ε of modulus N , this algorithm computes x∈A4 (N ) ε(x).

(1) [Factor N ] Compute the prime factorization pe11 · · · penn of N . (2) [Initialize] Set t = 1 and i = 0. (3) [Loop Over Prime Divisors] Set i = i + 1. If i > n, return t. Otherwise set p = pi and e = ei . (a) If p ≡ 3 (mod 4), return 0. (b) If p = 2 and e > 1, return 0. (c) If p = 2 and e = 1, go to step (3). (d) Compute a generator a ∈ (Z/pZ)∗ using Algorithm 4.4. (e) Compute ω = a(p−1)/4 . (f) Use the Chinese Remainder Theorem to find x ∈ Z/N Z such that x ≡ a (mod p) and x ≡ 1 (mod N/pe ). r−1 (g) Set x = xp . (h) Set s = ε(x). (i) If s = 1, set t = 2t and go to step (3). (j) If s = −1, set t = −2t and go to step (3).

Proof. Note that ε(−x) = ε(x), since ε is even. By the Chinese Remainder Theorem, the set A4 (N ) is empty if and only if there is no square root of −1 modulo some prime power divisor of p. If A4 (N ) is empty, the algorithm correctly detects this fact in steps (3a)–(3b). Thus assume A4 (N ) is nonempty. For each prime power pei i that exactly divides N , let xi ∈ Z/N Z e be such that x2i = −1 and xi ≡ 1 (mod pj j ) for i 6= j. This is the value of x computed in steps (3d)–(3g) (as one sees using elementary number theory). The next key observation is that Y (ε(xi ) + ε(−xi )) = (6.3.1) i

X

x∈A4 (N )

ε(x),

100

6. Dimension Formulas

since by the Chinese Remainder Theorem the elements of A4 (N ) are in bijection with the choices for a square root of −1 modulo each prime power divisors of N . The observation (6.3.1) is a huge gain from an efficiency point of view—if N had r prime factors, then A4 (N ) would have size 2r , which could be prohibitive, where the product involves only r factors. To finish the proof, just note that steps (3h)–(3j) compute the local factors ε(xi ) + ε(−xi ) = 2ε(xi ), where again we use that ε is even. Note that a solution of x2 + 1 ≡ 0 (mod p) lifts uniquely to a solution mod pn for any n, because the kernel of the natural homomorphism (Z/pn Z)∗ → (Z/pZ)∗ is a group of p-power order. ¤

The algorithm for computing the sum over A3 (N ) is similar. For k ≥ 2, to compute dim Sk (N, ε), use the formula directly and the fact that dim M2−k (N, ε) = 0, unless ε = 1 and k = 2. To compute dim Mk (N, ε) for k ≥ 2, use the fact that the big formula at the beginning of this section is valid for any integer k to replace k by 2 − k and that dim Sk (N, ε) = 0 for k ≤ 0 to rewrite the formula as dim Mk (N, ε) = −(dim S2−k (N, ε) − dim Mk (N, ε)) ³1 − k 1 Y λ(p, N, vp (c)) · µ0 (N ) − · =− 12 2 p|N X X ε(x) + γ3 (2 − k) · + γ4 (2 − k) · x∈A4 (N )

x∈A3 (N )

´ ε(x) .

Note also that for k = 0, dim Ek (N, ε) = 1 if and only if ε is trivial and it equals 0 otherwise. We then also obtain dim Ek (N, ε) = dim Mk (N, ε) − dim Sk (N, ε). We can also compute dim Ek (N, ε) when k = 1 directly, since dim S2−1 (N, ε) = dim S1 (N, ε). The following table contains the dimension of Sk (N, ε) for some sample values of N and k. In each case, ε is the product of characters εp of maximal order corresponding to the prime power factors of N (i.e., the product of the generators of the group D(N, C∗ ) of Dirichlet characters of modulus N ).

101

6.3. Modular Forms with Character

N 1 10 11 100 389 1000 2007

dim S2 (N, ε) 0 0 0 13 0 148 222

dim S3 (N, ε) 0 1 1 0 64 0 0

dim S4 (N, ε) 0 0 0 43 0 448 670

dim S24 (N, ε) 2 0 0 343 0 3448 5150

Example 6.10. We compute the last line of the above table. First we create the character ε. sage: G = DirichletGroup(2007) sage: e = prod(G.gens(), G(1)) Next we compute the dimension of the four spaces. sage: 222 sage: 0 sage: 670 sage: 5150

dimension_cusp_forms(e,2) dimension_cusp_forms(e,3) dimension_cusp_forms(e,4) dimension_cusp_forms(e,24)

We can also compute dimensions of the corresponding spaces of Eisenstein series. sage: 4 sage: 0 sage: 4 sage: 4

dimension_eis(e,2) dimension_eis(e,3) dimension_eis(e,4) dimension_eis(e,24)

Remark 6.11. Cohen and Oesterl´e also give dimension formulas for spaces of half-integral weight modular forms, which we do not give in this chapter. Note that [CO77] does not contain any proofs that their claimed formulas are correct, but instead they say only that “Les formules qui les donnent sont connues de beaucoup de gens et il existe plusieurs m´ethodes permettant de les obtenir (th´eor`eme de Riemann-Roch, application des formules de trace

102

6. Dimension Formulas

donn´ees par Shimura).”2 Fortunately, in [Que06], Jordi Quer derives the (integral weight) formulas of [CO77] along with formulas for dimensions of spaces Sk (G) and Mk (G) for more general congruence subgroups. Let f be the conductor of a Dirichlet character ε of modulus N . Then the dimension of the new subspace of Mk (N, ε) is X dim Sk (N, ε)new = µ(N/M ) · dim Sk (M, ε′ ), M such that f |M |N

where µ is as in the statement of Proposition 6.4, and ε′ is the restriction of ε mod M . Example 6.12. We compute the dimension of S2 (2007, ε)new for ε a quadratic character of modulus 2007. sage: G = DirichletGroup(2007, QQ) sage: e = prod(G.gens(), G(1)) sage: dimension_new_cusp_forms(e,2) 76

6.4. Exercises 6.1 Let µ0 and µ1 be as in this chapter. (a) Prove that µ0 (N ) = [SL2 (Z) : Γ0 (N )]. (b) Prove that for N ≥ 3, µ1 (N ) = [SL2 (Z) : Γ1 (N )]/2, so µ1 (N ) is the index of Γ1 (N ) · {±1} in PSL2 (Z) = SL2 (Z)/{±1}.

6.2 Use Proposition 6.4 to find a formula for dim Sk (SL2 (Z)). Verify that this formula is the same as the one in Corollary 2.16.

6.3 Suppose either that N = 1 or that N is prime and k = 2. Prove that Mk (Γ0 (N ))new = Mk (Γ0 (N )). 6.4 Fill in the details of the proof of Algorithm 6.9. 6.5 Implement a computer program to compute dim Sk (Γ0 (N )) as a function of k and N .

2The formulas that we give here are well known and there exist many methods to prove them, e.g., the Riemann-Roch theorem and applications of the trace formula of Shimura.

Chapter 7

Linear Algebra

This chapter is about several algorithms for matrix algebra over the rational numbers and cyclotomic fields. Algorithms for linear algebra over exact fields are necessary in order to implement the modular symbols algorithms that we will describe in Chapter 7. This chapter partly overlaps with [Coh93, Sections 2.1–2.4]. Note: We view all matrices as defining linear transformations by acting on row vectors from the right.

7.1. Echelon Forms of Matrices Definition 7.1 (Reduced Row Echelon Form). A matrix is in (reduced row) echelon form if each row in the matrix has more zeros at the beginning than the row above it, the first nonzero entry of every row is 1, and the first nonzero entry of any row is the only nonzero entry in its column. Given a matrix A, there is another matrix B such that B is obtained from A by left multiplication by an invertible matrix and B is in reduced row echelon form. This matrix B is called the echelon form of A. It is unique. A pivot column of A is a column of A such that the reduced row echelon form of A contains a leading 1. Example 7.2. The following matrix is not in reduced row echelon form:   14 2 7 228 −224  0 0 3 78 −70  . 0 0 0 −405 381 103

104

7. Linear Algebra

The reduced row echelon form of the above matrix is   1 71 0 0 − 1174 945  152   0 0 1 0 135  . 0 0 0 1 − 127 135

Notice that the entries of the reduced row echelon form can be rationals with large denominators even though the entries of the original matrix A are integers. Another example is the simple looking matrix   −9 6 7 3 1 0 0 0  −10 3 8 2 0 1 0 0     3 −6 2 8 0 0 1 0  −8 −6 −8 6 0 0 0 1 whose echelon form  1  0    0 0

is 0 1 0 0

0 0 1 0

42 0 1025 716 0 3075 83 0 − 1025 184 1 1025

92 − 1025 641 − 3075 133 1025 159 − 1025

1 25 2 − 75 1 25 2 25

9 − 205 7 − 615 23 − 410 9 410



  . 

A basic fact is that two matrices A and B have the same reduced row echelon form if and only if there is an invertible matrix E such that EA = B. Also, many standard operations in linear algebra, e.g., computation of the kernel of a linear map, intersection of subspaces, membership checking, etc., can be encoded as a question about computing the echelon form of a matrix. The following standard algorithm computes the echelon form of a matrix. Algorithm 7.3 (Gauss Elimination). Given an m×n matrix A over a field, the algorithm outputs the reduced row echelon form of A. Write ai,j for the i, j entry of A, where 0 ≤ i ≤ m − 1 and 0 ≤ j ≤ n − 1.

(1) [Initialize] Set k = 0. (2) [Clear Each Column] For each column c = 0, 1, . . . , n − 1, clear the cth column as follows: (a) [First Nonzero] Find the smallest r such that ar,c 6= 0, or if there is no such r, go to the next column. 1 times row r. (b) [Rescale] Replace row r of A by ar,c (c) [Swap] Swap row r with row k. (d) [Clear] For each i = 0, . . . , m − 1 with i 6= k, if ai,c 6= 0, add −ai,c times row k of A to row i to clear the leading entry of the ith row. (e) [Increment] Set k = k + 1.

This algorithm takes O(mn2 ) arithmetic operations in the base field, where A is an m × n matrix. If the base field is Q, the entries can become

7.2. Rational Reconstruction

105

huge and arithmetic operations are then very expensive. See Section 7.3 for ways to mitigate this problem. To conclude this section, we mention how to convert a few standard problems into questions about reduced row echelon forms of matrices. Note that one can also phrase some of these answers in terms of the echelon form, which might be easier to compute, or an LUP decomposition (lower triangular times upper triangular times permutation matrix), which the numerical analysts use. (1) Kernel of A: We explain how to compute the kernel of A acting on column vectors from the right (first transpose to obtain the kernel of A acting on row vectors). Since passing to the reduced row echelon form of A is the same as multiplying on the left by an invertible matrix, the kernel of the reduce row echelon form E of A is the same as the kernel of A. There is a basis vector of ker(E) that corresponds to each nonpivot column of E. That vector has a 1 at the nonpivot column, 0’s at all other nonpivot columns, and for each pivot column, the negative of the entry of A at the nonpivot column in the row with that pivot element. (2) Intersection of Subspaces: Suppose W1 and W2 are subspace of a finite-dimensional vector space V . Let A1 and A2 be matrices whose columns form a basis for W1 and W2 , respectively. Let A = [A1 |A2 ] be the augmented matrix formed from A1 and A2 . Let K be the kernel of the linear transformation defined by A. Then K is isomorphic to the desired intersection. To write down the intersection explicitly, suppose that dim(W1 ) ≤ dim(W2 ) and do the following: For each b in a basis for K, write down the linear combination of a basis for W1 obtained by taking the first dim(W1 ) entries of the vector b. The fact that b is in Ker(A) implies that the vector we just wrote down is also in W2 . This is because a linear relation X X ai w1,i + bj w2,j = 0, i.e., an element of that kernel, is the same as X X ai w1,i = −bj w2,j .

For more details, see [Coh93, Alg. 2.3.9].

7.2. Rational Reconstruction Rational reconstruction is a process that allows one to sometimes lift an integer modulo m uniquely to a bounded rational number.

106

7. Linear Algebra

Algorithm 7.4 (Rational Reconstruction). Given an integer a ≥ 0 and an integer m > 1, this algorithm computes the numerator n and denominator d of the unique rational number n/d, if it exists, with r m (7.2.1) |n|, d ≤ and n ≡ ad (mod m), 2 or it reports that there is no such number. (1) [Reduce mod m] Replace a with the least integer between 0 and m − 1 that is congruent to a modulo m. (2) [Trivial Cases] If a p = 0 or a = 1, return a. (3) [Initialize] Let b = m/2, u = m, v = a, and set U = (1, 0, u) and V = (0, 1, v). Use the notation Ui and Vi to refer to the ith entries of U, V , for i = 0, 1, 2. (4) [Iterate] Do the following as long as |V2 | > b: Set q = ⌊U2 /V2 ⌋, set T = U − qV , and set U = V and V = T . (5) [Numerator and Denominator] Set d = |V1 | and n = V2 . (6) [Good?] If d ≤ b and gcd(n, d) = 1, return n/d; otherwise report that there is no rational number as in (7.2.1). Algorithm 7.4 for rational reconstruction is described (with proof) in [Knu, pgs. 656–657] as the solution to Exercise 51 on page 379 in that book. See, in particular, the paragraph right in the middle of page 657, which describes the algorithm. Knuth attributes this rational reconstruction algorithm to Wang, Kornerup, and Gregory from around 1983. We now give an indication of why Algorithm 7.4 computes the rational reconstruction of a (mod m), leaving the precise details and uniqueness to [Knu, pgs. 656–657]. At each step in Algorithm 7.4, the 3-tuple V = (v0 , v1 , v2 ) satisfies (7.2.2)

m · v 0 + a · v 1 = v2 ,

and similarly for U . When computing the usual extended gcd, at the end v2 = gcd(a, m) and v0 , v1 give a representation of the v2 as a Z-linear combination of m and a. In Algorithm 7.4, we are instead interested in finding a rational number n/d such that n ≡ a · d (mod m). If we set n = v2 and d = v1 in (7.2.2) and rearrange, we obtain n = a · d + m · v0 . Thus at every step of the algorithm we find a rational number n/d such that n ≡ ad (mod m). The problem at intermediate steps is that, e.g., v0 could be 0, or n or d could be too large. Example 7.5. We compute an example using SAGE.

7.3. Echelon Forms over Q

sage: sage: sage: 210 sage: 7/13

107

p = 389 k = GF(p) a = k(7/13); a a.rational_reconstruction()

7.3. Echelon Forms over Q A difficulty with computation of the echelon form of a matrix over the rational numbers is that arithmetic with large rational numbers is timeconsuming; each addition potentially requires a gcd and numerous additions and multiplications of integers. Moreover, the entries of A during intermediate steps of Algorithm 7.3 can be huge even though the entries of A and the answer are small. For example, suppose A is an invertible square matrix. Then the echelon form of A is the identity matrix, but during intermediate steps the numbers involved could be quite large. One technique for mitigating this is to compute the echelon form using a multimodular method. If A is a matrix with rational entries, let H(A) be the height of A, which is the maximum of the absolute values of the numerators and denominators of all entries of A. If x, y are rational numbers and p is a prime, we write x ≡ y (mod p) to mean that the denominators of x and y are not divisible by p but the numerator of the rational number x − y (in reduced form) is divisible by p. For example, if x = 5/7 and y = 2/11, then x − y = 41/77, so x ≡ y (mod 41).

Algorithm 7.6 (Multimodular Echelon Form). Given an m × n matrix A with entries in Q, this algorithm computes the reduced row echelon form of A. (1) Rescale the input matrix A to have integer entries. This does not change the echelon form and makes reduction modulo many primes easier. We may thus assume A has integer entries. (2) Let c be a guess for the height of the echelon form. (3) List successive primes p1 , p2 , . . . such that the product of the pi is greater than n · c · H(A) + 1, where n is the number of columns of A. (4) Compute the echelon forms Bi of the reduction A (mod pi ) using, e.g., Algorithm 7.3 or any other echelon algorithm. (5) Discard any Bi whose pivot column list is not maximal among pivot lists of all Bj found so far. (The pivot list associated to Bi is the ordered list of integers k such that the kth column of Bj is a pivot

108

7. Linear Algebra

column. We mean maximal with respect to the following ordering on integer sequences: shorter integer sequences are smaller, and if two sequences have the same length, then order in reverse lexicographic order. Thus [1, 2] is smaller than [1, 2, 3], and [1, 2, 7] is smaller than [1, 2, 5]. Think of maximal as “optimal”, i.e., best possible pivot columns.) (6) Use the Chinese Remainder Theorem to find a matrix B with integer entries such that B ≡ Bi (mod pi ) for all pi . (7) Use Algorithm 7.4 to try to find a matrixpC whose coefficients are Q rational numbers n/r such that |n|, r ≤ M/2, where M = pi , and C ≡ Bi (mod pi ) for each prime p. If rational reconstruction fails, compute a few more echelon forms mod the next few primes (using the above steps) and attempt rational reconstruction again. Let E be the matrix over Q so obtained. (A trick here is to keep track of denominators found so far to avoid doing very many rational reconstructions.) (8) Compute the denominator d of E, i.e., the smallest positive integer such that dE has integer entries. If Y (7.3.1) H(dE) · H(A) · n < pi , then E is the reduced row echelon form of A. If not, repeat the above steps with a few more primes.

Proof. We prove that if (7.3.1) is satisfied, then the matrix E computed by the algorithm really is the reduced row echelon form R of A. First note that E is in reduced row echelon form since the set of pivot columns of all matrices Bi used to construct E are the same, so the pivot columns of E are the same as those of any Bi and all other entries in the Bi pivot columns are 0, so the other entries of E in the pivot columns are also 0. Recall from the end of Section 7.1 that a matrix whose columns are a basis for the kernel of A can be obtained from the reduced row echelon form R. Let K be the matrix whose columns are the vectors in the kernel algorithm applied to E, so EK = 0. Since the reduced row echelon form is obtained by left multiplying by an invertible matrix, for each i, there is an invertible matrix Vi mod pi such that A ≡ Vi Bi (mod pi ) so A · dK ≡ Vi Bi · dK ≡ Vi · dE · K ≡ 0 (mod pi ). Since dK and A are integer matrices, the Chinese remainder theorem implies that ³ Y ´ A · dK ≡ 0 mod pi .

The integer entries a of A · dK all satisfy |a| ≤ H(A) · H(dK) · n, where n is the number of columns of A. Since H(K) ≤ H(E), the bound (7.3.1)

7.3. Echelon Forms over Q

109

implies that A · dK = 0. Thus AK = 0, so Ker(E) ⊂ Ker(A). On the other hand, the rank of E equals the rank of each Bi (since the pivot columns are the same), so rank(E) = rank(Bi ) = rank(A (mod pi )) ≤ rank(A).

Thus dim(Ker(A)) ≤ dim(Ker(E)), and combining this with the bound obtained above, we see that Ker(E) = Ker(A). This implies that E is the reduced row echelon form of A, since two matrices have the same kernel if and only if they have the same reduced row echelon form (the echelon form is an invariant of the row space, and the kernel is the orthogonal complement of the row space). The reason for step (5) is that the matrices Bi need not be the reduction of R modulo pi , and indeed this reduction might not even be defined, e.g., if pi divides the denominator of some element of R, then this ¡ p 1reduction ¢ makes³ no sense. For example, set p = pi and suppose A = 0 0 . Then ´ R =

1 1/p 0 0

, which has no reduction modulo p; also, the reduction of A

modulo Bi is Bi = ( 00 10 ) (mod p), which is already in reduced row echelon form. However if we were to combine Bi with the echelon form of A modulo another prime, the result could never be lifted using rational reconstruction. Thus the reason we exclude all Bi with nonmaximal pivot column sequence is so that a rational reconstruction will exist. There are only finitely many primes that divide denominators of entries of R, so eventually all Bi will have maximal pivot column sequences, i.e., they are the reduction of the true reduced row echelon form R, so the algorithm terminates. ¤

Remark 7.7. Algorithm 7.6, with sparse matrices seems to work very well in practice. A simple but helpful modification to Algorithm 7.3 in the sparse case is to clear each column using a row with a minimal number of nonzero entries, so as to reduce the amount of “fill in” (denseness) of the matrix. There are much more sophisticated methods along these lines called “intelligent Gauss elimination”. (Cryptographers are interested in linear algebra mod p with huge sparse matrices, since they come up in attacks on the discrete log problem and integer factorization.) One can adapt Algorithm 7.6 to computation of echelon forms of matrices A over cyclotomic fields Q(ζn ). Assume A has denominator 1. Let p be a prime that splits completely in Q(ζn ). Compute the homomorphisms fi : Zp [ζn ] → Fp by finding the elements of order n in F∗p . Then compute the mod p matrix fi (A) for each i, and find its reduced row echelon form. Taken together, the maps fi together induce an isomorphism Ψ : Fp [X]/Φn (X) ∼ = Fdp , where Φn (X) is the nth cyclotomic polynomial and d is its degree. It is easy to compute Ψ(f (x)) by evaluating f (x) at each element of order n in Fp . To compute Ψ−1 , simply use linear algebra over Fp

110

7. Linear Algebra

to invert a matrix that represents Ψ. Use Ψ−1 to compute the reduced row echelon form of A (mod p), where (p) is the nonprime ideal in Z[ζn ] generated by p. Do this for several primes p, and use rational reconstruction on each coefficient of each power of ζn , to recover the echelon form of A.

7.4. Echelon Forms via Matrix Multiplication In this section we explain how to compute echelon forms using matrix multiplication. This is valuable because there are asymptotically fast, i.e., better than O(n3 ) field operations, algorithms for matrix multiplication, and implementations of linear algebra libraries often include highly optimized matrix multiplication algorithms. We only sketch the basic ideas behind these asymptotically fast algorithms (following [Ste]), since more detail would take us too far from modular forms. The naive algorithm for multiplying two m × m matrices requires O(m3 ) arithmetic operations in the base ring. In [Str69], Strassen described a clever algorithm that computes the product of two m × m matrices in O(mlog2 (7) ) = O(m2.807... ) arithmetic operations in the base ring. Because of numerical stability issues, Strassen’s algorithm is rarely used in numerical analysis. But for matrix arithmetic over exact base rings (e.g., the rational numbers, finite fields, etc.) it is of extreme importance. In [Str69], Strassen also sketched a new algorithm for computing the inverse of a square matrix using matrix multiplication. Using this algorithm, the number of operations to invert an m × m matrix is (roughly) the same as the number needed to multiply two m × m matrices. ¡ A B ¢Suppose the input n n matrix is 2 × 2 and we write it in block form as C D where A, B, C, D are all 2n−1 × 2n−1 matrices. Assume that any intermediate matrices below that we invert are invertible. Consider the augmented matrix ¶ µ A B I 0 . C D 0 I

Multiply the top row by A−1 to obtain µ ¶ I A−1 B A−1 0 , 0 I C D

and write E = A−1 B. Subtract C times the first row from the second row to get ¶ µ A−1 0 I E . 0 D − CE −CA−1 I

Set F = D − CE and multiply the bottom row by F −1 on the left to obtain ¶ µ I E A−1 0 . 0 I −F −1 CA−1 F −1

7.4. Echelon Forms via Matrix Multiplication

111

Set G = −F −1 CA−1 , and subtract E times the second from the first row to arrive at µ ¶ I 0 A−1 − EG −EF −1 . 0 I G F −1 The idea listed above can, with significant work, be extended to a general algorithm (as is done in [Ste06]). Next we very briefly sketch how to compute echelon forms of matrices using matrix multiplication and inversion. Its complexity is comparable to the complexity of matrix multiplication. As motivation, recall the standard algorithm from undergraduate linear algebra for inverting an invertible square matrix A: form the augmented matrix [A|I], and then compute the echelon form of this matrix, which is [I|A−1 ]. If T is the transformation matrix to echelon form, then T [A|I] = [I|T ], so T = A−1 . In particular, we could find the echelon form of [A|I] by multiplying on the left by A−1 . Likewise, for any matrix B with the same number of rows as A, we could find the echelon form of [A|B] by multiplying on the left by A−1 . Next we extend this idea to give an algorithm to compute echelon forms using only matrix multiplication (and echelon form modulo one prime). Algorithm 7.8 (Asymptotically Fast Echelon Form). Given a matrix A over the rational numbers (or a number field), this algorithm computes the echelon form of A. (1) [Find Pivots] Choose a random prime p (coprime to the denominator of any entry of A) and compute the echelon form of A (mod p), e.g., using Algorithm 7.3. Let c0 , . . . , cn−1 be the pivot columns of A (mod p). When computing the echelon form, save the positions r0 , . . . , rn−1 of the rows used to clear each column. (2) [Extract Submatrix] Extract the n × n submatrix B of A whose entries are Ari ,cj for 0 ≤ i, j ≤ n − 1. (3) [Compute Inverse] Compute the inverse B −1 of B. Note that B must be invertible since its reduction modulo p is invertible. (4) [Multiply] Let C be the matrix whose rows are the rows r0 , . . . , rn−1 of A. Compute E = B −1 C. If E is not in echelon form, go to step (1). (5) [Done?] Write down a matrix D whose columns are a basis for ker(E) as explained on page 105. Let F be the matrix whose rows are the rows of A other than rows r0 , . . . , rn−1 . Compute the product F D. If F D = 0, output E, which is the echelon form of A. If F D 6= 0, go to step (1) and run the whole algorithm again.

112

7. Linear Algebra

Proof. We prove both that the algorithm terminates and that when it terminates, the matrix E is the echelon form of A. First we prove that the algorithm terminates. Let E be the echelon form of A. By Exercise 7.3, for all but finitely many primes p (i.e., any prime where A (mod p) has the same rank as A) the echelon form of A (mod p) equals E (mod p). For any such prime p the pivot columns of E (mod p) are the pivot columns of E, so the algorithm will terminate for that choice of p. We next prove that when the algorithm terminates, E is the echelon form of A. By assumption, E is in echelon form and is obtained by multiplying C on the left by an invertible matrix, so E must be the echelon form of C. The rows of C are a subset of those of A, so the rows of E are a subset of the rows of the echelon form of A. Thus ker(A) ⊂ ker(E). To show that E equals the echelon form of A, we just need to verify that ker(E) ⊂ ker(A), i.e., that AD = 0, where D is as in step (5). Since E is the echelon form of C, we know that CD = 0. By step (5) we also know that F D = 0. Thus AD = 0, since the rows of A are the union of the rows of F and C. ¤ Example 7.9. Let A be the 4 × 8 matrix  −9 6 7 3  −10 3 8 2 A=  3 −6 2 8 −8 −6 −8 6

1 0 0 0

0 1 0 0

0 0 1 0

from Example 7.2.

 0 0   0  1

sage: M = MatrixSpace(QQ,4,8) sage: A = M([[-9,6,7,3,1,0,0,0],[-10,3,8,2,0,1,0,0], [3,-6,2,8,0,0,1,0],[-8,-6,-8,6,0,0,0,1]]) First choose the “random” prime p = 41, which does not divide any of the entries of A, and compute the echelon form of the reduction of A modulo 41. sage: A41 = MatrixSpace(GF(41),4,8)(A) sage: E41 = A41.echelon_form() The echelon form of A (mod 41) is  1 0 0 2 0  0 1 0 40 0   0 0 1 39 0 0 0 0 0 1

 20 33 18 30 7 1  . 19 13 17  31 0 37

113

7.4. Echelon Forms via Matrix Multiplication

Thus we take c0 = 0, c1 = 1, c2 = 2, and c3 = 4. Also ri = i for i = 0, 1, 2, 3. Next extract the submatrix B. sage: B = A.matrix_from_columns([0,1,2,4]) The submatrix B is

B

−1

Multiplying by A yields  E=B

 1 0  . 0  0

−9 6 7  −10 3 8 B=  3 −6 2 −8 −6 −8

The inverse of B is

−1



  A= 



  =  1 0 0 0

0 1 0 0

5 0 − 92 1 0 − 138 11 0 184 159 1 − 184

1 46 3 − 23 7 92 41 92

0 − 21 92 0 − 179 138 83 1 184 0 1025 184

9 − 184 11 − 276 17 − 368 45 368

5 0 − 92 1 0 − 138 11 0 184 1 − 159 184



  .  1 46 3 − 23 7 92 41 92

9 − 184 11 − 276 17 − 368

sage: E = B^(-1)*A

45 368



  . 

This is not the echelon form of A. Indeed, it is not even in echelon form, since the last row is not normalized so the leftmost nonzero entry is 1. We thus choose another random prime, say p = 43. The echelon form mod 43 has columns 0, 1, 2, 3 as pivot columns. We thus extract the matrix   −9 6 7 3  −10 3 8 2  . B=  3 −6 2 8  −8 −6 −8 6 sage: B = A.matrix_from_columns([0,1,2,3])

This matrix has inverse  B

−1

  = 

42 1025 716 3075 83 − 1025 184 1025

92 − 1025 641 − 3075 133 1025 159 − 1025

1 25 2 − 75 1 25 2 25

9 − 205 7 − 615 23 − 410 9 410



  . 

114

7. Linear Algebra

Finally, the echelon form of A is E = B −1 A. No further checking is needed since the product so obtained is in echelon form, and the matrix F of the last step of the algorithm has 0 rows. Remark 7.10. Above we have given only the barest sketch of asymptotically fast “block” algorithms for linear algebra. For optimized algorithms that work in the general case, please see the source code of [Ste06].

7.5. Decomposing Spaces under the Action of Matrix Efficiently solving the following problem is a crucial step in computing a basis of eigenforms for a space of modular forms (see Sections 3.7 and 9.3.2). Problem 7.11. Suppose T is an n × n matrix with entries in a field K (typically a number field or finite field) and that the minimal polynomial of T is square-free and has degree n. View T as acting on V = K n . Find a simple module decomposition W0 ⊕ · · · ⊕ Wm of V as a direct sum of simple K[T ]-modules. Equivalently, find an invertible matrix A such that A−1 T A is a block direct sum of matrices T0 , . . . , Tm such that the minimal polynomial of each Ti is irreducible. Remark 7.12. A generalization of Problem 7.11 to arbitrary matrices with entries in Q is finding the rational Jordan form (or rational canonical form, or Frobenius form) of T . This is like the usual Jordan form, but the resulting matrix is over Q and the summands of the matrix corresponding to eigenvalues are replaced by matrices whose minimal polynomials are the minimal polynomials (over Q) of the eigenvalues. The rational Jordan form was extensively studied by Giesbrecht in his Ph.D. thesis and many successive papers, where he analyzes the complexity of his algorithms and observes that the limiting step is factoring polynomials over K. The reason is that given a polynomial f ∈ K[x], one can easily write down a matrix T such that one can read off the factorization of f from the rational Jordan form of T (see also [Ste97]). 7.5.1. Characteristic Polynomials. The computation of characteristic polynomials of matrices is crucial to modular forms computations. There are many approaches to this problems: compute det(xI −A) symbolically (bad), compute the traces of the powers of A (bad), or compute the Hessenberg form modulo many primes and use CRT (bad; see for [Coh93, §2.2.4] the definition of Hessenberg form and the algorithm). A more sophisticated method is to compute the rational canonical form of A using Giesbrecht’s algorithm1 (see [GS02]), which involves computing Krylov subspaces (see Remark 7.13 below), and building up the whole space on which A acts. This 1Allan Steel also invented a similar algorithm.

115

7.5. Decomposing Spaces under the Action of Matrix

latter method is a generalization of Wiedemann’s algorithm for computing minimal polynomials (see Section 7.5.3), but with more structure to handle the case when the characteristic polynomial is not equal to the minimal polynomial. 7.5.2. Polynomial Factorization. Factorization of polynomials in Q[X] (or over number fields) is an important step in computing an explicit basis of Hecke eigenforms for spaces of modular forms. The best algorithm is the van Hoeij method [BHKS06], which uses the LLL lattice basis reduction algorithm [LLL82] in a novel way to solve the optimization problems that come up in trying to lift factorizations mod p to Z. It has been generalized by Belebas, van Hoeij, Kl¨ uners, and Steel to number fields. 7.5.3. Wiedemann’s Minimal Polynomial Algorithm. In this section we describe an algorithm due to Wiedemann for computing the minimal polynomial of an n × n matrix A over a field. Choose a random vector v and compute the iterates

(7.5.1)

v0 = v,

v1 = A(v),

v2 = A2 (v),

...,

v2n−1 = A2n−1 (v).

If f = xm + cm−1 xm−1 + · · · + c1 x + c0 is the minimal polynomial of A, then Am + cm−1 Am−1 + · · · + c0 In = 0,

where In is the n × n identity matrix. For any k ≥ 0, by multiplying both sides on the right by the vector Ak v, we see that Am+k v + cm−1 Am−1+k v + · · · + c0 Ak v = 0; hence vm+k + cm−1 vm−1+k + · · · + c0 vk = 0,

all k ≥ 0.

Wiedemann’s idea is to observe that any single component of the vectors v0 , . . . , v2n−1 satisfies the linear recurrence with coefficients 1, cm−1 , . . . , c0 . The Berlekamp-Massey algorithm (see Algorithm 7.14 below) was introduced in the 1960s in the context of coding theory to find the minimal polynomial of a linear recurrence sequence {ar }. The minimal polynomial of this linear recurrence is by definition the unique monic polynomial g, such that if {ar } satisfies a linear recurrence aj+k + bj−1 aj−1+k + · · · + b0 ak = 0 P i (for all k ≥ 0), then g divides the polynomial xj + j−1 i=0 bi x . If we apply Berlekamp-Massey to the top coordinates of the vi , we obtain a polynomial g0 , which divides f . We then apply it to the second to the top coordinates and find a polynomial g1 that divides f , etc. Taking the least common multiple of the first few gi , we find a divisor of the minimal polynomial of f . One can show that with “high probability” one quickly finds f , instead of just a proper divisor of f .

116

7. Linear Algebra

Remark 7.13. In the literature, techniques that involve iterating a vector as in (7.5.1) are often called Krylov methods. The subspace generated by the iterates of a vector under a matrix is called a Krylov subspace. Algorithm 7.14 (Berlekamp-Massey). Suppose a0 , . . . , a2n−1 are the first 2n terms of a sequence that satisfies a linear recurrence of degree at most n. This algorithm computes the minimal polynomial f of the sequence. P2n−1 (1) Let R0 = x2n , R1 = i=0 ai xi , V0 = 0, V1 = 1. (2) While deg(R1 ) ≥ n, do the following: (a) Compute Q and R such that R0 = QR1 + R. (b) Let (V0 , V1 , R0 , R1 ) = (V1 , V0 − QV1 , R1 , R). (3) Let d = max(deg(V1 ), 1 + deg(R1 )) and set P = xd V1 (1/x). (4) Let c be the leading coefficient of P and output f = P/c. The above description of Berlekamp-Massey is taken from [ADT04], which contains some additional ideas for improvements. Now suppose T is an n × n matrix as in Problem 7.11. We find the minimal polynomial of T by computing the minimal polynomial of T (mod p) using Wiedemann’s algorithm, for many primes p, and using the Chinese Remainder Theorem. (One has to bound the number of primes that must be considered; see, e.g., [Coh93].) One can also compute the characteristic polynomial of T directly from the Hessenberg form of T , which can be computed in O(n4 ) field operations, as described in [Coh93]. This is simple but slow. Also, the T we consider will often be sparse, and Wiedemann is particularly good when T is sparse. Example 7.15. We compute the minimal  3 0  0 0 A= −1 1/2

polynomial of  0 2  −1

using Wiedemann’s algorithm. Let v = (1, 0, 0)t . Then v = (1, 0, 0)t ,

Av = (3, 0, −1)t ,

A3 v = (27, −4, −8)t ,

A2 v = (9, −2, −2)t ,

A4 v = (81, −16, −21)t ,

A5 v = (243, −42, −68)t .

The linear recurrence sequence coming from the first entries is 1, 3, 9, 27, 81, 243. This sequence satisfies the linear recurrence ak+1 − 3ak = 0,

all k > 0,

so its minimal polynomial is x − 3. This implies that x − 3 divides the minimal polynomial of the matrix A. Next we use the sequence of second

7.5. Decomposing Spaces under the Action of Matrix

117

coordinates of the iterates of v, which is 0, 0, −2, −4, −16, −42. The recurrence that this sequence satisfies is slightly less obvious, so we apply the Berlekamp-Massey algorithm to find it, with n = 3. (1) We have R0 = x6 , R1 = −42x5 − 16x4 − 4x3 − 2x2 , V0 = 0, V1 = 1.

(2) (a) Dividing R0 by R1 , we find

µ ¶ µ ¶ 1 22 4 4 5 3 8 2 R0 = R1 − x + x − x + x . + 42 441 441 441 441 (b) The new V0 , V1 , R0 , R1 are V0 = 1, 4 1 , V1 = x − 42 441 R0 = −42x5 − 16x4 − 4x3 − 2x2 , 22 4 5 3 8 2 R1 = x − x + x . 441 441 441

Since deg(R1 ) ≥ n = 3, we do the above three steps again.

(3) We repeat the above three steps. (a) Dividing R0 by R1 , we find

¶ µ ¶ µ 1323 3 882 2 123921 9261 + x− x + x . R0 = R1 − 11 242 242 121 (b) The new V0 , V1 , R0 , R1 are: 1 4 x− , 42 441 441 2 2205 441 V1 = x + x+ , 22 484 121 5 3 8 2 22 4 x − x + x , R0 = 441 441 441 1323 3 882 2 x + x . R1 = 242 121 V0 =

118

7. Linear Algebra

(4) We have to repeat the steps yet again: 441 441 2 2205 x + x+ , 22 484 121 242 3 968 2 484 242 V1 = − x + x + x− , 1323 3969 3969 3969 1323 3 882 2 x + x , R0 = 242 121 484 2 x . R1 = 3969 V0 =

242 3 x + (5) We have d = 3, so P = − 3969

484 2 3969 x

+

968 3969 x

(6) Multiply through by −3969/242 and output

−

242 1323 .

x3 − 2x2 − 4x + 3 = (x − 3)(x2 + x − 1).

The minimal polynomial of T2 is (x − 3)(x2 + x − 1), since the minimal polynomial has degree at most 3 and is divisible by (x − 3)(x2 + x − 1). 7.5.4. p-adic Nullspace. We will use the following algorithm of Dixon [Dix82] to compute p-adic approximations to solutions of linear equations over Q. Rational reconstruction modulo pn then allows us to recover the corresponding solutions over Q. Algorithm 7.16 (p-adic Nullspace). Given a matrix A with integer entries and nonzero kernel, this algorithm computes a nonzero element of ker(A) using successive p-adic approximations. (1) [Prime] Choose a random prime p. (2) [Echelon] Compute the echelon form of A modulo p. (3) [Done?] If A has full rank modulo p, it has full rank, so we terminate the algorithm. (4) [Setup] Let b0 = 0. (5) [Iterate] For each m = 0, 1, 2, . . . , k, use the echelon form of A modulo p to find a vector ym with integer entries such that Aym ≡ bm (mod p), and then set bm+1 =

bm − Aym . p

(If m = 0, choose ym 6= 0.) (6) [p-adic Solution] Let x = y0 + y1 p + y2 p2 + y3 p3 + · · · + yk pk . (7) [Lift] Use rational reconstruction (Algorithm 7.4) to find a vector z with rational entries such that z ≡ x (mod pk+1 ), if such a vector exists. If the vector does not exist, increase k or use a different p. Otherwise, output z.

119

7.6. Exercises

Proof. We verify the case k = 2 only. We have Ay0 = 0 (mod p) and Ay1 = − Ayp 0 (mod p). Thus Ay0 + pAy1 ≡ Ay0 + (−Ay0 ) (mod p2 ). ¤ 7.5.5. Decomposition Using Kernels. We now know enough to give an algorithm to solve Problem 7.11. Algorithm 7.17 (Decomposition Using Kernels). Given an n × n matrix T over a field K as in Problem 7.11, this algorithm computes the decomposition of V as a direct sum of simple K[T ] modules. (1) [Minimal Polynomial] Compute the minimal polynomial f of T , e.g., using the multimodular Wiedemann algorithm. (2) [Factorization] Factor f using the algorithm in Section 7.5.2. (3) [Compute Kernels] For each irreducible factor gi of f , compute the following. (a) Compute the matrix Ai = gi (T ). (b) Compute Wi = ker(Ai ), e.g., using Algorithm 7.16. L (4) [Output Answer] Then V = Wi .

Remark 7.18. As mentioned in Remark 7.12, if one can compute such L decompositions V = Wi , then one can easily factor polynomials f ; hence the difficulty of polynomial factorization is a lower bound on the complexity of writing V as a direct sum of simples.

7.6. Exercises 7.1 Given a subspace W of k n , where k is a field and n ≥ 0 is an integer, give an algorithm to find a matrix A such that W = Ker(A). 7.2 If rref(A) denotes the row reduced echelon form of A and p is a prime not dividing any denominator of any entry of A or of rref(A), is rref(A (mod p)) = rref(A) (mod p)? 7.3 Let A be a matrix with entries in Q. Prove that for all but finitely many primes p we have rref(A (mod p)) = rref(A) (mod p). 7.4 Let



 1 2 3 A =  4 5 6 . 7 8 9

(a) Compute the echelon form of A using each of Algorithm 7.3, Algorithm 7.6, and Algorithm 7.8. (b) Compute the kernel of A.

120

7. Linear Algebra

(c) Find the characteristic polynomial of A using the algorithm of Section 7.5.3. 7.5 The notion of echelon form extends to matrices whose entries come from certain rings other than fields, e.g., Euclidean domains. In the case of matrices over Z we define a matrix to be in echelon form (or Hermite normal form) if it satisfies • aij = 0, for i > j, • aii ≥ 0, • aij < aii for all j < i (unless aii = 0, in which case all aij = 0). There are algorithms for computing with finitely generated modules over Z that are analogous to the ones in this chapter for vector spaces, which depend on computation  of Hermite  forms.   1 2 3 1 2 3 (a) Show that the Hermite form of 4 5 6 is 0 3 6. 7 8 9 0 0 0 (b) Describe an algorithm for transforming an n×n matrix A with integer entries into Hermite form using row operations and the Euclidean algorithm.

Chapter 8

General Modular Symbols

In this chapter we explain how to generalize the notion of modular symbols given in Chapter 3 to higher weight and more general level. We define Hecke operators on them and their relation to modular forms via the integration pairing. We omit many difficult proofs that modular symbols have certain properties and instead focus on how to compute with modular symbols. For more details see the references given in this section (especially [Mer94]) and [Wie05]. Modular symbols are a formalism that make it elementary to compute with homology or cohomology related to certain Kuga-Sato varieties (these are E ×X · · · ×X E, where X is a modular curve and E is the universal elliptic curve over it). It is not necessary to know anything about these Kuga-Sato varieties in order to compute with modular symbols. This chapter is about spaces of modular symbols and how to compute with them. It is by far the most important chapter in this book. The algorithms that build on the theory in this chapter are central to all the computations we will do later in the book. This chapter closely follows Lo¨ıc Merel’s paper [Mer94]. First we define modular symbols of weight k ≥ 2. Then we define the corresponding Manin symbols and state a theorem of Merel-Shokurov, which gives all relations between Manin symbols. (The proof of the Merel-Shokurov theorem is beyond the scope of this book but is presented nicely in [Wie05].) Next we describe how the Hecke operators act on both modular and Manin symbols

121

122

8. General Modular Symbols

and how to compute trace and inclusion maps between spaces of modular symbols of different levels. Not only are modular symbols useful for computation, but they have been used to prove theoretical results about modular forms. For example, certain technical calculations with modular symbols are used in Lo¨ıc Merel’s proof of the uniform boundedness conjecture for torsion points on elliptic curves over number fields (modular symbols are used to understand linear independence of Hecke operators). Another example is [Gri05], which distills hypotheses about Kato’s Euler system in K2 of modular curves to a simple formula involving modular symbols (when the hypotheses are satisfied, one obtains a lower bound on the Shafarevich-Tate group of an elliptic curve).

8.1. Modular Symbols We recall from Chapter 3 the free abelian group M2 of modular symbols. We view these as elements of the relative homology of the extended upper half plane h∗ = h ∪ P1 (Q) relative to the cusps. The group M2 is the free abelian group on symbols {α, β} with α, β ∈ P1 (Q) = Q ∪ {∞}

modulo the relations {α, β} + {β, γ} + {γ, α} = 0,

for all α, β, γ ∈ P1 (Q), and all torsion. More precisely, M2 = (F/R)/(F/R)tor , where F is the free abelian group on all pairs (α, β) and R is the subgroup generated by all elements of the form (α, β) + (β, γ) + (γ, α). Note that M2 is a huge free abelian group of countable rank. For any integer n ≥ 0, let Z[X, Y ]n be the abelian group of homogeneous polynomials of degree n in two variables X, Y . Remark 8.1. Note that Z[X, Y ]n is isomorphic to Symn (Z × Z) as a group, but certain natural actions are different. In [Mer94], Merel uses the notation Zn [X, Y ] for what we denote by Z[X, Y ]n . Now fix an integer k ≥ 2. Set Mk = Z[X, Y ]k−2 ⊗Z M2 , which is a torsion-free abelian group whose elements are sums of expressions of the form X i Y k−2−i ⊗ {α, β}. For example, X 3 ⊗ {0, 1/2} − 17XY 2 ⊗ {∞, 1/7} ∈ M5 .

123

8.1. Modular Symbols

Fix a finite index subgroup ¡ G¢ of SL2 (Z). Define a left action of G on Z[X, Y ]k−2 as follows. If g = ac db ∈ G and P (X, Y ) ∈ Z[X, Y ]k−2 , let (gP )(X, Y ) = P (dX − bY, −cX + aY ).

Note that if we think of z = (X, Y ) as a column vector, then (gP )(z) = P (g −1 z),

¡ ¢ since g −1 = −cd −ba . The reason for the inverse is so that this is a left action instead of a right action, e.g., if g, h ∈ G, then ((gh)P )(z) = P ((gh)−1 z) = P (h−1 g −1 z) = (hP )(g −1 z) = (g(hP ))(z). Recall that we let G act on the left on M2 by g{α, β} = {g(α), g(β)}, where G acts via linear fractional transformations, so if g = g(α) =

aα + b . cα + d

¡a b¢ c d

For example, useful special cases to remember are that if g = a b and g(∞) = . d c (Here we view ∞ as 1/0 in order to describe the action.) g(0) =

, then

¡a b¢ c d

, then

We now combine these two actions to obtain a left action of G on Mk , which is given by g(P ⊗ {α, β}) = (gP ) ⊗ {g(α), g(β)}. For example, ¾ ¶ ½ µ 2 5 1 2 3 3 (X ⊗ {0, 1/2}) = (−3X − 2Y ) ⊗ − , − −2 −3 3 8

½

2 5 = (−27X − 54X Y − 36XY − 8Y ) ⊗ − , − 3 8 3

2

2

3

¾

.

We will often write P (X, Y ){α, β} for P (X, Y ) ⊗ {α, β}. Definition 8.2 (Modular Symbols). Let k ≥ 2 be an integer and let G be a finite index subgroup of SL2 (Z). The space Mk (G) of weight k modular symbols for G is the quotient of Mk by all relations gx−x for x ∈ Mk , g ∈ G, and by any torsion. Note that Mk is a torsion-free abelian group, and it is a nontrivial fact that Mk has finite rank. We denote modular symbols for G in exactly the same way we denote elements of Mk ; the group G will be clear from context.

124

8. General Modular Symbols

The space of modular symbols over a ring R is Mk (G; R) = Mk (G) ⊗Z R.

8.2. Manin Symbols Let G be a finite index subgroup of SL2 (Z) and k ≥ 2 an integer. Just as in Chapter 3 it is possible to compute Mk (G) using a computer, despite that, as defined above, Mk (G) is the quotient of one infinitely generated abelian group by another one. This section is about Manin symbols, which are a distinguished subset of Mk (G) that lead to a finite presentation for Mk (G). Formulas written in terms of Manin symbols are frequently much easier to compute using a computer than formulas in terms of modular symbols. Suppose P ∈ Z[X, Y ]k−2 and g ∈ SL2 (Z). Then the Manin symbol associated to this pair of elements is [P, g] = g(P {0, ∞}) ∈ Mk (G). Notice that if Gg = Gh, then [P, g] = [P, h], since the symbol g(P {0, ∞}) is invariant by the action of G on the left (by definition, since it is a modular symbol for G). Thus for a right coset Gg it makes sense to write [P, Gg] for the symbol [P, h] for any h ∈ Gg. Since G has finite index in SL2 (Z), the abelian group generated by Manin symbols is of finite rank, generated by © k−2−i i ª [X Y , Ggj ] : i = 0, . . . , k − 2 and j = 0, . . . , r , where g0 , . . . , gr run through representatives for the right cosets G\ SL2 (Z).

We next show that every modular symbol can be written as a Z-linear combination of Manin symbols, so they generate Mk (G). Proposition 8.3. The Manin symbols generate Mk (G). Proof. The proof if very similar to that of Proposition 3.11 except we introduce an extra twist to deal with the polynomial part. Suppose that we are given a modular symbol P {α, β} and wish to represent it as a sum of Manin symbols. Because P {a/b, c/d} = P {a/b, 0} + P {0, c/d}, it suffices to write P {0, a/b} in terms of Manin symbols. Let 0=

p0 p1 p2 0 p−1 1 p0 pr a p−2 = , = , = , , , ..., = q−2 1 q−1 0 1 q0 q1 q2 qr b

denote the continued fraction convergents of the rational number a/b. Then pj qj−1 − pj−1 qj = (−1)j−1

for − 1 ≤ j ≤ r.

125

8.2. Manin Symbols

If we let gj =

µ

(−1)j−1 pj (−1)j−1 qj

¶ pj−1 , then gj ∈ SL2 (Z) and qj−1

¾ r ½ X pj−1 pj P {0, a/b} = P , qj−1 qj =

=

j=−1 r X

j=−1 r X

gj ((gj−1 P ){0, ∞})

[gj−1 P, gj ].

j=−1

Since gj ∈ SL2 (Z) and P has integer coefficients, the polynomial gj−1 P also has integer coefficients, so we introduce no denominators. ¤ Now that we know the Manin symbols generate Mk (G), we next consider the relations between Manin symbols. Fortunately, the answer is fairly simple (though the proof is not). Let ¶ ¶ µ ¶ µ µ −1 0 0 −1 0 −1 . , J= , τ= σ= 0 −1 1 −1 1 0 Define a right action of SL2 (Z) on Manin symbols as follows. If h ∈ SL2 (Z), let [P, g]h = [h−1 P, gh]. This is a right action because both P 7→ h−1 P and g 7→ gh are right actions. Theorem 8.4. If x is a Manin symbol, then (8.2.1)

x + xσ = 0,

(8.2.2)

x + xτ + xτ 2 = 0,

(8.2.3)

x − xJ = 0.

Moreover, these are all the relations between Manin symbols, in the sense that the space Mk (G) of modular symbols is isomorphic to the quotient of the free abelian group on the finitely many symbols [X i Y k−2−i , Gg] (for i = 0, . . . , k − 2 and Gg ∈ G\ SL2 (Z)) by the above relations and any torsion. Proof. First we prove that the Manin symbols satisfy the above relations. We follow Merel’s proof (see [Mer94, §1.2]). Note that σ(0) = σ 2 (∞) = ∞

and

τ (1) = τ 2 (0) = ∞.

126

8. General Modular Symbols

Writing x = [P, g], we have [P, g] + [P, g]σ = [P, g] + [σ −1 P, gσ] = g(P {0, ∞}) + gσ(σ −1 P {0, ∞})

= (gP ){g(0), g(∞)} + (gσ)(σ −1 P ){gσ(0), gσ(∞)}

= (gP ){g(0), g(∞)} + (gP ){g(∞), g(0)} = (gP ){g(0), g(∞)} + {g(∞), g(0)})

= 0. Also,

[P, g] + [P, g]τ + [P, g]τ 2 = [P, g] + [τ −1 P, gτ ] + [τ −2 P, gτ 2 ] = g(P {0, ∞}) + gτ (τ −1 P {0, ∞}) + gτ 2 (τ −2 P {0, ∞})

= (gP ){g(0), g(∞)} + (gP ){gτ (0), gτ (∞)} + (gP ){gτ 2 (0), τ 2 (∞)} = (gP ){g(0), g(∞)} + (gP ){g(1), g(0)}) + (gP ){g(∞), g(1)} ¡ ¢ = (gP ) {g(0), g(∞)} + {g(∞), g(1)} + {g(1), g(0)}

= 0. Finally,

[P, g] + [P, g]J = g(P {0, ∞}) − gJ(J −1 P {gJ(0), gJ(∞)} = (gP ){g(0), g(∞)} − (gP ){g(0), g(∞)}

= 0,

where we use that J acts trivially via linear fractional transformations. This proves that the listed relations are all satisfied. That the listed relations are all relations is more difficult to prove. One approach is to show (as in [Mer94, §1.3]) that the quotient of Manin symbols by the above relations and torsion is isomorphic to a space of Shokurov symbols, which is in turn isomorphic to Mk (G). A much better approach is to apply some results from group cohomology, as in [Wie05]. ¤ If G is a finite index subgroup and we have an algorithm to enumerate the right cosets G\ SL2 (Z) and to decide which coset an arbitrary element of SL2 (Z) belongs to, then Theorem 8.4 and the algorithms of Chapter 7 yield an algorithm to compute Mk (G; Q). Note that if J ∈ G, then the relation x − xJ = 0 is automatic. Remark 8.5. The matrices σ and τ do not commute, so in computing Mk (G; Q), one cannot first quotient out by the two-term σ relations, then quotient out only the remaining free generators by the τ relations, and get the right answer in general.

127

8.2. Manin Symbols

8.2.1. Coset Representatives and Manin Symbols. The following is analogous to Proposition 3.10: Proposition 8.6. The right cosets Γ1 (N )\ SL2 (Z) are in bijection with pairs (c, ¡ a d) ¢ where c, d ∈ Z/N Z and gcd(c, d, N ) = 1. The coset containing a matrix b corresponds to (c, d). c d

Proof. This proof is copied from [Cre92, pg. 203], except in that paper Cremona works with the analogue ³ of´ Γ1 (N ) in PSL2 (Z), so his result is slightly different. Suppose γi = acii dbii ∈ SL2 (Z), for i = 1, 2. We have ¶ ¶ µ ¶µ µ a1 d2 − b1 c2 ∗ d2 −b2 a1 b1 −1 , = γ1 γ2 = c1 d2 − d1 c2 a2 d1 − b2 c1 −c2 a2 c1 d 1

which is in Γ1 (N ) if and only if

c1 d2 − d1 c2 ≡ 0 (mod N )

(8.2.4) and

a2 d1 − b2 c1 ≡ a1 d2 − b1 c2 ≡ 1 (mod N ).

(8.2.5)

Since the γi have determinant 1, if (c1 , d1 ) = (c2 , d2 ) (mod N ), then the congruences (8.2.4)–(8.2.5) hold. Conversely, if (8.2.4)–(8.2.5) hold, then c 2 ≡ a2 d 1 c 2 − b 2 c 1 c 2

≡ a2 d 2 c 1 − b 2 c 2 c 1

≡ c1

and likewise

since d1 c2 ≡ d2 c1

(mod N )

since a2 d2 − b2 c2 = 1,

d2 ≡ a2 d1 d2 − b2 c1 d2 ≡ a2 d1 d2 − b2 d1 c2 ≡ d1

(mod N ). ¤

Thus we may view weight k Manin symbols for Γ1 (N ) as triples of integers (i, c, d), where 0 ≤ i ≤ k − 2 and c, d ∈ Z/N Z with ¡gcd(c,¢d, N ) = 1. Here (i, c, d) corresponds to the Manin symbol [X i Y k−2−i , ca′ db′ ], where c′ and d′ are congruent to c, d (mod N ). The relations of Theorem 8.4 become

(i, c, d)

+

(i, c, d) + (−1)i (k − 2 − i, d, −c) = 0, µ ¶ k−2−i X k−2−i (−1)k−2 (−1)j (j, d, −c − d) j j=0

+

k−2−i

(−1)

i X j=0

µ ¶ i (−1) (k − 2 − i + j, −c − d, c) = 0, j j

(i, c, d) − (−1)k−2 (i, −c, −d) = 0.

Recall that Proposition 3.10 gives a similar description for Γ0 (N ).

128

8. General Modular Symbols

8.2.2. Modular Symbols with Character. Suppose G = Γ1 (N ). Define an action of diamond-bracket operators hdio, with gcd(d, N ) = 1 on Manin symbols as follows: hdi([P, (u, v)]) = [P, (du, dv)] . Let ε : (Z/N Z)∗ → Q(ζ)∗

be a Dirichlet character, where ζ is an nth root of unity and n is the order of ε. Let Mk (N, ε) be the quotient of Mk (Γ1 (N ); Z[ζ]) by the relations (given in terms of Manin symbols) hdix − ε(d)x = 0,

for all x ∈ Mk (Γ1 (N ); Z[ζ]), d ∈ (Z/N Z)∗ , and by any Z-torsion. Thus Mk (N, ε) is a Z[ε]-module that has no torsion when viewed as a Z-module.

8.3. Hecke Operators Suppose Γ is a subgroup of SL2 (Z) of level N that contains Γ1 (N ). Just as for modular forms, there is a commutative Hecke algebra T = Z[T1 , T2 , . . .], which is the subring of End(Mk (Γ)) generated by all Hecke operators. Let ¶¾ ¶ ¾ ½µ ½µ p 0 1 r , : r = 0, 1, . . . , p − 1 ∪ Rp = 0 1 0 p ¡ ¢ where we omit p0 10 if p | N . Then the Hecke operator Tp on Mk (Γ) is given by X Tp (x) = gx. g∈Rp

Notice when p ∤ N that Tp is defined by summing over p + 1 matrices that correspond to the p + 1 subgroups of Z × Z of index p. This is exactly how we defined Tp on modular forms in Definition 2.26.

8.3.1. General Definition of Hecke Operators. Let Γ be a finite index subgroup of SL2 (Z) and suppose ∆ ⊂ GL2 (Q) is a set such that Γ∆ = ∆Γ = ∆ and Γ\∆ is finite. For example, ∆ = Γ satisfies this condition. Also, if Γ = Γ1 (N ), then for any positive integer n, the set ¶ ¾ ¶ µ ¶ µ ½µ 1 ∗ a b a b (mod N ) ≡ ∈ Mat2 (Z) : ad − bc = n, ∆n = 0 n c d c d also satisfies this condition, as we will now prove.

129

8.3. Hecke Operators

Lemma 8.7. We have Γ1 (N ) · ∆n = ∆n · Γ1 (N ) = ∆n and ∆n = ´

³

[ a,b

¶ µ a b , Γ1 (N ) · σa 0 n/a

0 (mod N ), the union is disjoint and 1 ≤ a ≤ n with where σa ≡ 1/a 0 a a | n, gcd(a, N ) = 1, and 0 ≤ b < n/a. In particular, the set of cosets Γ1 (N )\∆n is finite.

Proof. (This is Lemma 1 of [Mer94, §2.3].) If γ ∈ Γ1 (N ) and δ ∈ ∆n , then ¶ ¶ µ ¶ µ ¶ µ ¶ µ µ 1 ∗ 1 ∗ 1 ∗ 1 ∗ 1 ∗ (mod N ). ≡ · ≡ · 0 n 0 1 0 n 0 n 0 1 Thus Γ1 (N )∆n ⊂ ∆n , and since Γ1 (N ) is a group, Γ1 (N )∆n = ∆n ; likewise ∆n Γ1 (N ) = ∆n . For the coset decomposition, we first prove the statement for N = 1, i.e., for Γ1 (N ) = SL2 (Z). If A is an arbitrary element of Mat2 (Z) with determinant n, then using row operators on the left with determinant 1, i.e., left by elements of SL2 (Z), we can transform A into the ´ ³ multiplication a b form 0 n/a , with 1 ≤ a ≤ n and 0 ≤ b < n. (Just imagine applying the Euclidean algorithm to the two entries in the first column of A. Then a is the gcd of the two entries in the first column, and the lower left entry is 0. Next subtract n/a from b until 0 ≤ b < n/a.) Next suppose N is arbitrary. Let g1 , . . . , gr be such that g1 Γ1 (N ) ∪ · · · ∪ gr Γ1 (N ) = SL2 (Z) is a disjoint union. If A ∈ ∆n is arbitrary, ´ then as we showed above, there is ³ a b some γ ∈ SL2 (Z), so that γ · A = 0 n/a , with 1 ≤ a ≤ n and 0 ≤ b < n/a, and a | n. Write γ = gi · α, with α ∈ Γ1 (N ). Then ¶ ¶ µ µ 1 ∗ a b −1 (mod N ). ≡ α · A = gi · 0 n 0 n/a It follows that gi−1

≡

µ

¶ ¶−1 µ ¶ µ 1/a ∗ a b 1 ∗ ≡ · 0 a 0 n/a 0 n

(mod N ).

Since ( 10 11 ) ∈ Γ1 (N ) and gcd(a, N ) = 1, there is γ ′ ∈ Γ1 (N ) such that ¶ µ 1/a 0 ′ −1 (mod N ). γ gi ≡ 0 a

130

8. General Modular Symbols

′ −1 We ³may then ´ choose σa = γ gi . Thus every A ∈ ∆n is of the form b , with γ ∈ Γ1 (N ) and a, b suitably bounded. This proves the γσa a0 n/a second claim. ¤

Let any element δ = Mk ⊗ Q by

¡a b¢ c d

∈ GL2 (Q) act on the left on modular symbols

δ(P {α, β}) = P (dX − bY, −cX + aY ){δ(α), δ(β)}. (Until now ¢ had only defined an action of SL2 (Z) on modular symbols.) ¡ a bwe For g = c d ∈ GL2 (Q), let ¶ µ d −b = det(g) · g −1 . (8.3.1) g˜ = −c a Note that g˜ ˜ = g. Also, δP (X, Y ) = (P ◦ g˜)(X, Y ), where we set g˜(X, Y ) = (dX − bY, −cX + aY ). Suppose Γ and ∆ are as above. Fix a finite set R of representatives for Γ\∆. Let T∆ : Mk (Γ) → Mk (Γ) be the linear map T∆ (x) =

X

δx.

δ∈R

This map is well defined because if γ ∈ Γ and x ∈ Mk (Γ), then X X X X δx, δγx = γδ ′ x = δ′x = δ∈R

certain δ ′

certain δ ′

δ∈R

where we have used that ∆Γ = Γ∆, and Γ acts trivially on Mk (Γ). Let Γ = Γ1 (N ) and ∆ = ∆n . Then the nth Hecke operator Tn is T∆n , and by Lemma 8.7, ¶ X µa b · x, σa Tn (x) = 0 n/a a,b

where a, b are as in Lemma 8.7. Given this definition, we can compute the Hecke operators on Mk (Γ1 (N )) as follows. Write x as a modular symbol P {α, β}, compute Tn (x) as a modular symbol, and then convert to Manin symbols using continued fractions expansions. This is extremely inefficient; fortunately Lo¨ıc Merel (following [Maz73]) found a much better way, which we now describe (see [Mer94]).

131

8.3. Hecke Operators

8.3.2. Hecke Operators on Manin Symbols. If S is a subset of GL2 (Q), let S˜ = {˜ g : g ∈ S}, where g˜ is as in (8.3.1). Also, for any ring R and any subset S ⊂ Mat2 (Z), let R[S] denote the free R-module with basis the elements of S, so the elements of R[S] are the finite R-linear combinations of the elements of S. One of the main theorems of [Mer94] is that for any Γ, ∆Psatisfying the condition at the beginning of Section 8.3.1, if we can find uM M ∈ C[Mat2 (Z)] and a map ˜ SL2 (Z) → SL2 (Z) φ:∆ that satisfies certain conditions, then for any Manin symbol [P, g] ∈ Mk (Γ), we have X ˜ · P, φ(gM )]. T∆ ([P, g]) = u M [M ˜ SL2 (Z) with M ∈SL2 (Z) gM ∈∆

The paper [Mer94] contains many examples of φ and that satisfy all of the conditions.

P

uM M ∈ C[Mat2 (Z)]

When Γ = Γ1 (N ), the complicated list of conditions becomes simpler. Let Mat2 (Z)n be the set of 2 × 2 matrices with determinant n. An element X h= uM [M ] ∈ C[Mat2 (Z)n ] satisfies condition Cn if for every K ∈ Mat2 (Z)n / SL2 (Z), we have that X (8.3.2) uM ([M ∞] − [M 0]) = [∞] − [0] ∈ C[P 1 (Q)]. M ∈K

If h satisfies condition Cn , then for any Manin symbol [P, g] ∈ Mk (Γ1 (N )), Merel proves that X (8.3.3) Tn ([P, (u, v)]) = uM [P (aX + bY, cX + dY ), (u, v)M ]. M

Here (u, v) ∈ (Z/N Z)2 corresponds via Proposition 8.6 to a coset of Γ1 (N ) in SL2 (Z), and if (u′ , v ′ ) = (u, v)M ∈ (Z/N Z)2 and gcd(u′ , v ′ , N ) 6= 1, then we omit the corresponding summand. For example, we will now check directly that the element ¶¸ ¶¸ ·µ ¶¸ ·µ ¶¸ ·µ ·µ 1 0 2 1 1 0 2 0 + + + h2 = 1 2 0 1 0 2 0 1

132

8. General Modular Symbols

satisfies condition C2 . We have, as in the proof of Lemma 8.7 (but using elementary column operations), that ¶ ¾ ½µ a 0 SL2 (Z) : a = 1, 2 and 0 ≤ b < 2/a Mat2 (Z)2 / SL2 (Z) = b 2/a ¶ ¾ ¶ µ ¶ µ ½µ 2 0 1 0 1 0 SL2 (Z) . SL2 (Z), SL2 (Z), = 0 1 1 2 0 2 To verify condition C2 , we consider in turn each of the three elements of Mat2 (Z)2 / SL2 (Z) and check that (8.3.2) holds. We have that ¶ ¶ µ µ 1 0 1 0 SL2 (Z), ∈ 0 2 0 2 ¶ ¶ µ ¶ µ µ 1 0 1 0 2 1 SL2 (Z), ∈ , 1 2 1 2 0 1

and

µ

¶ ¶ µ 2 0 2 0 SL2 (Z). ∈ 0 1 0 1

Thus if K = ( 10 02 ) SL2 (Z), the left sum of (8.3.2) is [( 10 02 ) (∞)]−[( 10 02 ) (0)] = [∞] − [0], as required. If K = ( 11 02 ) SL2 (Z), then the left side of (8.3.2) is [( 20 11 ) (∞)] − [( 20 11 ) (0)] + [( 11 02 ) (∞)] − [( 11 02 ) (0)] = [∞] − [1] + [1] − [0] = [∞] − [0].

Finally, for K = ( 20 01 ) SL2 (Z) we also have [( 20 01 ) (∞)]−[( 20 01 ) (0)] = [∞]−[0], as required. Thus by (8.3.3) we can compute T2 on any Manin symbol, by summing over the action of the four matrices ( 20 01 ) , ( 10 02 ) , ( 20 11 ) , ( 11 02 ). Proposition 8.8 (Merel). The element X ·µa b ¶¸ ∈ Z[Mat2 (Z)n ] c d a>b≥0 d>c≥0 ad−bc=n

satisfies condition Cn . Merel’s two-page proof of Proposition 8.8 is fairly elementary. Remark 8.9. In [Cre97a, §2.4], Cremona discusses the work of Merel and Mazur on Heilbronn matrices in the special cases Γ = Γ0 (N ) and weight 2. He gives a simple proof that the action of Tp on Manin symbols can be computed by summing the action of some set Rp of matrices of determinant p. He then describes the set Rp and gives an efficient continued fractions algorithm for computing it (but he does not prove that his Rp satisfy Merel’s hypothesis).

8.4. Cuspidal Modular Symbols

133

8.3.3. Remarks on Complexity. Merel gives another family Sn of matrices that satisfy condition Cn , and he proves that as n → ∞, 12 log(2) #Sn ∼ · σ1 (n) log(n), π2 where σ1 (n) is the sum of the divisors of n. Thus for a fixed space Mk (Γ) of modular symbols, one can compute Tn using O(σ1 (n) log(n)) arithmetic operations. Note that we have fixed Mk (Γ), so we ignore the linear algebra involved in computation of a presentation; also, adding elements takes a bounded number of field operations when the space is fixed. Thus, using Manin symbols the complexity of computing Tp , for p prime, is O((p + 1) log(p)) field operations, which is exponential in the number of digits of p. 8.3.4. Basmaji’s Trick. There is a trick of Basmaji (see [Bas96]) for computing a matrix of Tn on Mk (Γ), when n is very large, and it is more efficient than one might naively expect. Basmaji’s trick does not improve the big-oh complexity for a fixed space, but it does improve the complexity by a constant factor of the dimension of Mk (Γ; Q). Suppose we are interested in computing the matrix for Tn for some massive integer n and that Mk (Γ; Q) has large dimension. The trick is as follows. Choose a list x1 = [P1 , g1 ], . . . , xr = [Pr , gr ] ∈ V = Mk (Γ; Q)

of Manin symbols such that the map Ψ : T → V r given by t 7→ (tx1 , . . . , txr )

is injective. In practice, it is often possible to do this with r “very small”. Also, we emphasize that V r is a Q-vector space of dimension r · dim(V ).

Next find Hecke operators Ti , with i small, whose images form a basis for the image of Ψ. Now with the above data precomputed, which only required working with Hecke operators Ti for small i, we are ready to compute Tn with n huge. Compute yi = Tn (xi ), for each i = 1, . . . , r, which we can compute using Heilbronn matrices since each xi = [Pi , gi ] is a Manin symbol. We thus obtain Ψ(Tn ) ∈ V r . Since we have precomputedP Hecke operators Tj r such that Ψ(Tj ) generate V , we can findPaj such that aj Ψ(Tj ) = Ψ(Tn ). Then since Ψ is injective, we have Tn = aj Tj , which gives the full matrix of Tn on Mk (Γ; Q).

8.4. Cuspidal Modular Symbols Let B be the free abelian group on symbols {α}, for α ∈ P1 (Q), and set Bk = Z[X, Y ]k−2 ⊗ B.

Define a left action of SL2 (Z) on Bk by

g(P {α}) = (gP ){g(α)},

134

8. General Modular Symbols

for g ∈ SL2 (Z). For any finite index subgroup Γ ⊂ SL2 (Z), let Bk (Γ) be the quotient of Bk by the relations x − gx for all g ∈ Γ and by any torsion. Thus Bk (Γ) is a torsion-free abelian group. The boundary map is the map given by extending the map

b : Mk (Γ) → Bk (Γ)

b(P {α, β}) = P {β} − P {α}

linearly. The space Sk (Γ) of cuspidal modular symbols is the kernel Sk (Γ) = ker(Mk (Γ) → Bk (Γ)),

so we have an exact sequence

0 → Sk (Γ) → Mk (Γ) → Bk (Γ).

One can prove that when k > 2, this sequence is exact on the right. Next we give a presentation of Bk (Γ) in terms of “boundary Manin symbols”. 8.4.1. Boundary Manin Symbols. We give an explicit description of the boundary map in terms of Manin symbols for Γ = Γ1 (N ), then describe an efficient way to compute the boundary map. Let R be the equivalence relation on Γ\Q2 given by ¡ ¢ k u [Γ λu λv ] ∼ sign(λ) [Γ ( v )],

for any λ ∈ Q∗ . Denote by Bk (Γ) the finite-dimensional Q-vector space with basis the equivalence classes (Γ\Q2 )/R. The following two propositions are proved in [Mer94]. Proposition 8.10. The map µ : Bk (Γ) → Bk (Γ),

P

nuo v

· µ ¶¸ u 7→ P (u, v) Γ v

is well defined and injective. Here u and v are assumed coprime. Thus the kernel of δ : Sk (Γ) → Bk (Γ) is the same as the kernel of µ ◦ δ. ¡ ¢ Proposition 8.11. Let P ∈ Vk−2 and g = ac db ∈ SL2 (Z). We have ¡ ¢ µ ◦ δ([P, (c, d)]) = P (1, 0)[Γ ( ac )] − P (0, 1)[Γ db ].

We next describe how to explicitly compute µ ◦ δ : Mk (N, ε) → Bk (N, ε) by generalizing the ¡algorithm of [Cre97a, §2.2]. To compute the image of ¢ a b [P, ¡(c, ¢d)], with g = c d ∈ SL2 (Z), we must compute the class of [( ac )] and of [ db ]. Instead of finding a canonical form for cusps, we use a quick test for equivalence modulo scalars. In the following algorithm, by the ith symbol we

135

8.4. Cuspidal Modular Symbols

mean the ith basis vector for a basis of Bk (N, ε). This basis is constructed as the algorithm is called successively. We first give the algorithm, and then prove the facts used by the algorithm in testing equivalence. Algorithm 8.12 (Cusp Representation). Given a boundary Manin symbol [( uv )], this algorithm outputs an integer i and a scalar α such that [( uv )] is equivalent to α times the ith symbol found so far. (We call this algorithm repeatedly and maintain a running list of cusps seen so far.) (1) Use Proposition 3.21 to check whether or not [( uv )] is equivalent, modulo scalars, to any cusp found. If so, return the representative, its index, and the scalar. If not, record ( uv ) in the representative list. (2) Using Proposition 8.16, check whether or not [( uv )] is forced to equal 0 by the relations. If it does not equal 0, return its position in the list and the scalar 1. If it equals 0, return the scalar 0 and the position 1; keep ( uv ) in the list, and record that it is equivalent to 0. The case considered in Cremona’s book [Cre97a] only involve the trivial character, so no cusp classes are forced to vanish. Cremona gives the following two criteria for equivalence. Proposition 8.13 (Cremona). Consider ( uvii ), i = 1, 2, with ui , vi integers such that gcd(ui , vi ) = 1 for each i. (1) There exists g ∈ Γ0 (N ) such that g ( uv11 ) = ( uv22 ) if and only if s1 v2 ≡ s2 v1

(mod gcd(v1 v2 , N )), where sj satisfies uj sj ≡ 1 (mod vj ).

(2) There exists g ∈ Γ1 (N ) such that g ( uv11 ) = ( uv22 ) if and only if v2 ≡ v 1

(mod N ) and u2 ≡ u1

(mod gcd(v1 , N )).

Proof. The first statement is [Cre97a, Prop. 2.2.3], and the second is [Cre92, Lem. 3.2]. ¤ Algorithm 8.14 (Explicit Cusp Equivalence). Suppose ( uv11 ) and ( uv22 ) are equivalent modulo Γ0 (N ). This algorithm computes a matrix g ∈ Γ0 (N ) such that g ( uv11 ) = ( uv22 ). (1) Let s1 , s2 , r1 , r2 be solutions to s1 u1 − r1 v1 = 1 and s2 u2 − r2 v2 = 1. (2) Find integers x0 and y0 such that x0 v1 v2 + y0 N = 1. (3) Let x = −x0 (s1 v2 − s2 v1 )/(v1 v2 , N ) and s′1 = s1 + xv1 . ³ u r ´−1 1 1 (4) Output g = ( uv22 sr22 ) · v1 s′1 , which sends ( uv11 ) to ( uv22 ).

Proof. See the proof of [Cre97a, Prop. 8.13].

¤

136

8. General Modular Symbols

The ε relations can make the situation more complicated, since it is possible that ε(α) 6= ε(β) but ·µ ¶¸ ·µ ¶¸ · µ ¶¸ · µ ¶¸ u u u u . = ε(β) = γβ = γα ε(α) v v v v One way out of this difficulty is to construct the cusp classes for Γ1 (N ), and then quotient out by the additional ε relations using Gaussian elimination. This is far too inefficient to be useful in practice because the number of Γ1 (N ) cusp classes can be unreasonably large. Instead, we give a quick test to determine whether or not a cusp vanishes modulo the ε-relations. Lemma 8.15. Suppose α and α′ are integers such that gcd(α, α′ , N ) = 1. Then there exist integers β and β ′ , congruent to α and α′ modulo N , respectively, such that gcd(β, β ′ ) = 1. Proof. By Exercise 8.2 the map SL2 (Z) → SL2 (Z/N Z) is surjective. By the Euclidean algorithm, there¡ exist ¢integers x, y and z such that xα+yα′ +zN = 1. Consider the matrix αy −x ∈ SL2 (Z/N Z) and take β, β ′ to be the α′ bottom row of a lift of this matrix to SL2 (Z). ¤ Proposition 8.16. Let N be a positive integer and ε a Dirichlet character of modulus N . Suppose ( uv ) is a cusp with u and v coprime. Then ( uv ) vanishes modulo the relations [γ ( uv )] = ε(γ) [( uv )] ,

all γ ∈ Γ0 (N ),

if and only if there exists α ∈ (Z/N Z)∗ , with ε(α) 6= 1, such that v ≡ αv

u ≡ αu

(mod N ), (mod gcd(v, N )).

Proof. First suppose such an α exists. By Lemma 8.15³there´ exists β, β ′ ∈ Z lifting α, α−1 such that gcd(β, β ′ ) = 1. The cusp ββu has coprime ′v coordinates ³ ´so, by Proposition 8.13 and our congruence conditions on α, the cusps ββu and ( uv ) are equivalent by an element of Γ1 (N ). This implies ′ h³ v ´i h³ ´i βu βu u that = [( )]. Since = ε(α) [( uv )] and ε(α) 6= 1, we have ′ v βv β′v [( uv )] = 0. Conversely, suppose [( uv )] = 0. Because all relations are two-term relations and the Γ1 (N )-relations identify Γ1 (N )-orbits, there must exists α and β with · µ ¶¸ · µ ¶¸ u u γα = γβ and ε(α) 6= ε(β). v v

137

8.5. Pairing Modular Symbols and Modular Forms

Indeed, if this did not occur, then we could mod out by the ε relations by writing each [γα ( uv )] in terms of [( uv )], and there would be no further relations left to kill [( uv )]. Next observe that · µ ¶¸ · µ ¶¸ u u γβ −1 α = γβ −1 γα v v · µ ¶¸ · µ ¶¸ ·µ ¶¸ u u u −1 −1 = ε(β ) γα . = ε(β ) γβ = v v v

Applying Proposition 8.13 and noting that ε(β −1 α) 6= 1 shows that β −1 α satisfies the properties of the “α” in the statement of the proposition. ¤

We enumerate the possible α appearing in Proposition 8.16 as follows. Let g = (v, N ) and list the α = v · Ng · a + 1, for a = 0, . . . , g − 1, such that ε(α) 6= 0.

8.5. Pairing Modular Symbols and Modular Forms In this section we define a pairing between modular symbols and modular forms that the Hecke operators respect. We also define an involution on modular symbols and study its relationship with the pairing. This pairing is crucial in much that follows, because it gives rise to period maps from modular symbols to certain complex vector spaces. Fix an integer weight k ≥ 2 and a finite index subgroup Γ of SL2 (Z). Let Mk (Γ) denote the space of holomorphic modular forms of weight k for Γ, and let Sk (Γ) denote its cuspidal subspace. Following [Mer94, §1.5], let S k (Γ) = {f : f ∈ Sk (Γ)}

(8.5.1)

denote the space of antiholomorphic cusp forms. Here f is the function on h∗ given by f (z) = f (z). Define a pairing (Sk (Γ) ⊕ S k (Γ)) × Mk (Γ) → C

(8.5.2) by letting

h(f1 , f2 ), P {α, β}i =

Z

β

α

f1 (z)P (z, 1)dz +

Z

β

f2 (z)P (z, 1)dz

α

and extending linearly. Here the integral is a complex path integral along a simple path from α to β in h (so, e.g., write z(t) = x(t) + iy(t), where (x(t), y(t)) traces out the path and consider two real integrals). Proposition 8.17. The integration pairing is well defined, i.e., if we replace P {α, β} by an equivalent modular symbol (equivalent modulo the left action of Γ), then the integral is the same.

138

8. General Modular Symbols

Proof. This follows from the change of variables formulas for integration and the fact that f1 ∈ Sk (Γ) and f2 ∈ S k (Γ). For example, if k = 2, g ∈ Γ and f ∈ Sk (Γ), then hf, g{α, β}i = hf, {g(α), g(β)}i Z g(β) f (z)dz = = =

Z

g(α) β

f (g(z))dg(z)

α Z β α

f (z)dz = hf, {α, β}i,

where f (g(z))dg(z) = f (z)dz because f is a weight 2 modular form. For the case of arbitrary weight k > 2, see Exercise 8.4. ¤ The integration pairing is very relevant to thePstudy of special values of L-functions. The L-function of a cusp form f = an q n ∈ Sk (Γ1 (N )) is Z ∞ dt (8.5.3) L(f, s) = (2π)s Γ(s)−1 f (it)ts t 0 ∞ X an (8.5.4) for Re(s) ≫ 0. = ns n=1

The equality of the integral and the Dirichlet series follows by switching the order of summation and integration, which is justified using standard estimates on |an | (see, e.g., [Kna92, Section VIII.5]).

For each integer j with 1 ≤ j ≤ k − 1, we have, setting s = j and making the change of variables t 7→ −it in (8.5.3), that E (−2πi)j D (8.5.5) L(f, j) = · f, X j−1 Y k−2−(j−1) {0, ∞} . (j − 1)! The integers j as above are called critical integers. When f is an eigenform, they have deep conjectural significance (see [BK90, Sch90]). One can approximate L(f, j) to any desired precision by computing the above pairing explicitly using the method described in Chapter 10. Alternatively, [Dok04] contains methods for computing special values of very general L-functions, which can be used for approximating L(f, s) for arbitrary s, in addition to just the critical integers 1, . . . , k − 1. Theorem 8.18 (Shokoruv). The pairing h· , ·i : (Sk (Γ) ⊕ S k (Γ)) × Sk (Γ, C) → C

is a nondegenerate pairing of complex vector spaces.

Proof. This is [Sho80b, Thm. 0.2] and [Mer94, §1.5].

¤

8.5. Pairing Modular Symbols and Modular Forms

139

Corollary 8.19. We have dimC Sk (Γ, C) = 2 dimC S2 (Γ). The pairing is also compatible with Hecke operators. Before proving this, we define an action of Hecke operators on Mk (Γ1 (N )) and on S k (Γ1 (N )). The definition is similar to the one we gave in Section 2.4 for modular forms of level 1. For a positive integer n, let Rn be a set ¡ of¢ coset representatives for Γ1 (N )\∆n from Lemma 8.7. For any γ = ac db ∈ GL2 (Q) and f ∈ Mk (Γ1 (N )) set f [γ]k = det(γ)k−1 (cz + d)−k f (γ(z)). Also, for f ∈ S k (Γ1 (N )), set ′

f [γ]k = det(γ)k−1 (cz + d)−k f (γ(z)).

Then for f ∈ Mk (Γ1 (N )), Tn (f ) =

X

f [γ]k

γ∈Rn

and for f ∈ S k (Γ1 (N )),

Tn (f ) =

X

′

f [γ]k .

γ∈Rn

This agrees with the definition from Section 2.4 when N = 1. Remark 8.20. If Γ is an arbitrary finite index subgroup of SL2 (Z), then we can define operators T∆ on Mk (Γ) for any ∆ with ∆Γ = Γ∆ = ∆ and Γ\∆ finite. For concreteness we do not do the general case here or in the theorem below, but the proof is exactly the same (see [Mer94, §1.5]). Finally we prove the promised Hecke compatibility of the pairing. This proof should convince you that the definition of modular symbols is sensible, in that they are natural objects to integrate against modular forms. Theorem 8.21. If f = (f1 , f2 ) ∈ Sk (Γ1 (N )) ⊕ S k (Γ1 (N )) and x ∈ Mk (Γ1 (N )), then for any n, hTn (f ), xi = hf, Tn (x)i. Proof. We follow [Mer94, §2.1] (but with more details) and will only prove the theorem when f = f1 ∈ Sk (Γ1 (N )), the proof in the general case being the same. ¡ ¢ Let α, β ∈ P1 (Q), P ∈ Z[X, Y ]k−2 , and for g = ac db ∈ GL2 (Q), set j(g, z) = cz + d.

140

8. General Modular Symbols

Let n be any positive integer, and let Rn be a set of coset representatives for Γ1 (N )\∆n from Lemma 8.7. We have hTn (f ), P {α, β}i = =

Z

β

Tn (f )P (z, 1)dz

α

XZ

β

det(δ)k−1 f (δ(z))j(δ, z)−k P (z, 1)dz.

δ∈R α

Now for each summand corresponding to the δ ∈ R, make the change of variables u = δz. Thus we make #R change of variables. Also, we will use the notation ¶ µ d −b = det(g) · g −1 g˜ = −c a for g ∈ GL2 (Q). We have hTn (f ), P {α, β}i = X Z δ(β)

det(δ)k−1 f (u)j(δ, δ −1 (u))−k P (δ −1 (u), 1)d(δ −1 (u)).

δ∈R δ(α)

˜ We have δ −1 (u) = δ(u), since a linear fractional transformation is unchanged by a nonzero rescaling of a matrix that induces it. Thus by the quotient rule, using that δ˜ has determinant det(δ), we see that det(δ)du ˜ d(δ −1 (u)) = d(δ(u)) = . ˜ u)2 j(δ, We next show that (8.5.6)

˜ u)k det(δ)−k P (δ(u), ˜ j(δ, δ −1 (u))−k P (δ −1 (u), 1) = j(δ, 1).

˜ From the definitions, and again using that δ −1 (u) = δ(u), we see that j(δ, δ −1 (u)) =

det(δ) , ˜ u) j(δ,

which proves that (8.5.6) holds. Thus hTn (f ), P {α, β}i = X Z δ(β) det(δ)du ˜ u)k det(δ)−k P (δ(u), ˜ det(δ)k−1 f (u)j(δ, 1) . ˜ u)2 j(δ, δ∈R δ(α) Next we use that ˜ u)k−2 P (δ(u), ˜ (δP )(u, 1) = j(δ, 1).

8.5. Pairing Modular Symbols and Modular Forms

141

To see this, note that P (X, Y ) = P (X/Y, 1) · Y k−2 . Using this we see that ˜ (δP )(X, Y ) = (P ◦ δ)(X, Y) µ µ ¶ ¶µ ¶k−2 X X ˜ =P δ · Y k−2 . +a ,1 −c · Y Y

Now substituting (u, 1) for (X, 1), we see that ˜ (δP )(u, 1) = P (δ(u), 1) · (−cu + a)k−2 , as required. Thus finally hTn (f ), P {α, β}i = =

XZ

δ(β)

δ∈R δ(α) X Z δ(β)

δ∈R δ(α)

˜ u)k−2 P (δ(u), ˜ f (u)j(δ, 1)du f (u) · ((δP )(u, 1))du

= hf, Tn (P {α, β})i. ¤ that Γ is a finite index subgroup of SL2 (Z) such that if η = ¢ ¡ −1Suppose 0 , then 01 ηΓη = Γ. For example, Γ = Γ1 (N ) satisfies this condition. There is an involution ι∗ on Mk (Γ) given by (8.5.7)

ι∗ (P (X, Y ){α, β}) = −P (X, −Y ){−α, −β},

which we call the star involution. On Manin symbols, ι∗ is (8.5.8)

ι∗ [P, (u, v)] = −[P (−X, Y ), (−u, v)].

Let Sk (Γ)+ be the +1 eigenspace for ι∗ on Sk (Γ), and let Sk (Γ)− be the −1 eigenspace. There is also a map ι on modular forms, which is adjoint to ι∗ . Remark 8.22. Notice the minus sign in front of −P (X, −Y ){−α, −β} in (8.5.7). This sign is missing in [Cre97a], which is a potential source of confusion (this is not a mistake, but a different choice of convention). We now state the final result about the pairing, which explains how modular symbols and modular forms are related. Theorem 8.23. The integration pairing h· , ·i induces nondegenerate Hecke compatible bilinear pairings Sk (Γ)+ × Sk (Γ) → C

and

Sk (Γ)− × S k (Γ) → C.

142

8. General Modular Symbols

Remark 8.24. We make some remarks about computing the boundary map of Section 8.4.1 when working in the ±1 quotient. Let s be a sign, either +1 or −1. To compute Sk (N, ε)s , it is necessary to replace Bk (N, ε) by its quotient modulo the additional relations [( −uv )] = s [( uv )] for all cusps ( uv ). Algorithm 8.12 can be modified to deal with this situation as follows. Given a cusp x = ( uv ), proceed as in Algorithm 8.12 and check if either ( uv ) or ( −uv ) is equivalent (modulo scalars) to any cusp seen so far. If not, use the following trick to determine whether the ε and s-relations kill the class of ( uv ): use the unmodified Algorithm 8.12 to compute the scalars α1 , α2 and indices i1 , i2 associated to ( uv ) and ( −uv ), respectively. The s-relation kills the class of ( uv ) if and only if i1 = i2 but α1 6= sα2 .

8.6. Degeneracy Maps In this section, we describe natural maps between spaces of modular symbols with character of different levels. We consider spaces with character, since they are so important in applications. Fix a positive integer N and a Dirichlet character ε : (Z/N Z)∗ → C∗ . Let M be a positive divisor of N that is divisible by the conductor of ε, in the sense that ε factors through (Z/M Z)∗ via the natural map (Z/N Z)∗ → (Z/M Z)∗ composed with some uniquely defined character ε′ : (Z/M Z)∗ → C∗ . For any positive divisor t of N/M , let T = ( 10 0t ) and fix a choice Dt = {T γi : i = 1, . . . , n} of coset representatives for Γ0 (N )\T Γ0 (M ). Remark 8.25. Note that [Mer94, §2.6] contains a typo: The quotient “Γ1 (N )\Γ1 (M )T ” should be replaced by “Γ1 (N )\T Γ1 (M )”. Proposition 8.26. For each divisor t of N/M there are well-defined linear maps ¶ µ t 0 ′ −1 x, αt : Mk (N, ε) → Mk (M, ε ), αt (x) = (tT )x = 0 1 X βt : Mk (M, ε′ ) → Mk (N, ε), βt (x) = ε′ (γi )−1 T γi x. T γi ∈Dt

Furthermore, αt ◦ βt is multiplication by tk−2 · [Γ0 (M ) : Γ0 (N )]. Proof. To show that ¡ α ¢ t is well defined, we must show that for each x ∈ Mk (N, ε) and γ = ac db ∈ Γ0 (N ), we have αt (γx − ε(γ)x) = 0 ∈ Mk (M, ε′ ).

We have αt (γx) =

¶ ¶ µ ¶µ ¶ µ t 0 t 0 a tb t 0 ′ x, x = ε (a) γx = 0 1 0 1 c/t d 0 1

µ

143

8.6. Degeneracy Maps

so αt (γx − ε(γ)x) = ε′ (a)αt (x) − ε(γ)αt (x) = 0.

We next verify that βt is well defined. Suppose that x ∈ Mk (M, ε′ ) and γ ∈ Γ0 (M ); then ε′ (γ)−1 γx = x, so X ε′ (γi )−1 T γi ε′ (γ)−1 γx βt (x) = T γi ∈Dt

=

X

ε′ (γi γ)−1 T γi γx.

T γi γ∈Dt

This computation shows that the definition of βt does not depend on the choice Dt of coset representatives. To finish the proof that βt is well defined, we must show that, for γ ∈ Γ0 (M ), we have βt (γx) = ε′ (γ)βt (x) so that βt respects the relations that define Mk (M, ε). Using that βt does not depend on the choice of coset representative, we find that for γ ∈ Γ0 (M ), X βt (γx) = ε′ (γi )−1 T γi γx T γi ∈Dt

= =

X

ε′ (γi γ −1 )−1 T γi γ −1 γx

T γi γ −1 ∈Dt ε′ (γ)βt (x).

To compute αt ◦ βt , we use that #Dt = [Γ0 (N ) : Γ0 (M )]:   X ε′ (γi )−1 T γi x αt (βt (x)) = αt  =

X

T γi

ε′ (γi )−1 (tT −1 )T γi x

T γi

= tk−2

X

ε′ (γi )−1 γi x

T γi

= tk−2

X

x

T γi

= tk−2 · [Γ0 (N ) : Γ0 (M )] · x. The scalar factor of tk−2 appears instead of t, because t is acting on x as an element of GL2 (Q) and not as an an element of Q. ¤ Definition 8.27 (New and Old Modular Symbols). The space Mk (N, ε)new of new modular symbols is the intersection of the kernels of the αt as t runs through all positive divisors of N/M and M runs through positive divisors of M strictly less than N and divisible by the conductor of ε. The subspace

144

8. General Modular Symbols

Mk (N, ε)old of old modular symbols is the subspace generated by the images of the βt where t runs through all positive divisors of N/M and M runs through positive divisors of M strictly less than N and divisible by the conductor of ε. The new and old subspaces of cuspidal modular symbols are the intersections of the above spaces with Sk (N, ε). Example 8.28. The new and old subspaces need not be disjoint, as the following example illustrates! (This contradicts [Mer94, pg. 80].) Consider, for example, the case N = 6, k = 2, and trivial character. The spaces M2 (Γ0 (2)) and M2 (Γ0 (3)) are each of dimension 1, and each is generated by the modular symbol {∞, 0}. The space M2 (Γ0 (6)) is of dimension 3 and is generated by the three modular symbols {∞, 0}, {−1/4, 0}, and {−1/2, −1/3}. The space generated by the two images of M2 (Γ0 (2)) under the two degeneracy maps has dimension 2, and likewise for M2 (Γ0 (3)). Together these images generate M2 (Γ0 (6)), so M2 (Γ0 (6)) is equal to its old subspace. However, the new subspace is nontrivial because the two degeneracy maps M2 (Γ0 (6)) → M2 (Γ0 (2)) are equal, as are the two degeneracy maps M2 (Γ0 (6)) → M2 (Γ0 (3)). In particular, the intersection of the kernels of the degeneracy maps has dimension at least 1 (in fact, it equals 1). We verify some of the above claims using SAGE. sage: M = ModularSymbols(Gamma0(6)); M Modular Symbols space of dimension 3 for Gamma_0(6) of weight 2 with sign 0 over Rational Field sage: M.new_subspace() Modular Symbols subspace of dimension 1 of Modular Symbols space of dimension 3 for Gamma_0(6) of weight 2 with sign 0 over Rational Field sage: M.old_subspace() Modular Symbols subspace of dimension 3 of Modular Symbols space of dimension 3 for Gamma_0(6) of weight 2 with sign 0 over Rational Field

8.7. Explicitly Computing Mk (Γ0 (N )) In this section we explicitly compute Mk (Γ0 (N )) for various k and N . We represent Manin symbols for Γ0 (N ) as triples of integers (i, u, v), where (u, v) ∈ P1 (Z/N Z), and (i, u, v) corresponds to [X i Y k−2−i , (u, v)] in the usual notation. Also, recall from Proposition 3.10 that (u, v) corresponds to

8.7. Explicitly Computing Mk (Γ0 (N ))

145

¡ ¢ the right coset of Γ0 (N ) that contains a matrix ac db with (u, v) ≡ (c, d) as elements of P1 (Z/N Z), i.e., up to rescaling by an element of (Z/N Z)∗ . 8.7.1. Computing P1 (Z/N Z). In this section we give an algorithm to compute a canonical representative for each element of P1 (Z/N Z). This algorithm is extremely important because modular symbols implementations use it a huge number of times. A more naive approach would be to store all pairs (u, v) ∈ (Z/N Z)2 and a fixed reduced representative, but this wastes a huge amount of memory. For example, if N = 105 , we would store an array of 2 · 105 · 105 = 20 billion integers.

Another approach to enumerating P1 (Z/N Z) is described at the end of [Cre97a, §2.2]. It uses the fact that is easy to test whether two pairs (u0 , v0 ), (u1 , v1 ) define the same element of P1 (Z/N Z); they do if and only if we have equality of cross terms u0 v1 = v0 u1 (mod N ) (see [Cre97a, Prop. 2.2.1]). So we consider the 0-based list of elements (8.7.1)

(1, 0), (1, 1), . . . , (1, N − 1), (0, 1)

concatenated with the list of nonequivalent elements (d, a) for d | N and a = 1, . . . , N − 1, checking each time we add a new element to our list (of (d, a)) whether we have already seen it. Given a random pair (u, v), the problem is then to find the index of the element of our list of the equivalent representative in P1 (Z/N Z). We use the following algorithm, which finds a canonical representative for each element of P1 (Z/N Z). Given an arbitrary (u, v), we first find the canonical equivalent elements (u′ , v ′ ). If u′ = 1, then the index is v ′ . If u′ 6= 1, we find the corresponding element in an explicit sorted list, e.g., using binary search. In the following algorithm, a (mod N ) denotes the residue of a modulo N that satisfies 0 ≤ a < N . Note that we never create and store the list (8.7.1) itself in memory. Algorithm 8.29 (Reduction in P1 (Z/N Z) to Canonical Form). Given u and v and a positive integer N , this algorithm outputs a pair u0 , v0 such that (u, v) ≡ (u0 , v0 ) as elements of P1 (Z/N Z) and s ∈ Z such that (u, v) = (su0 , sv0 ) (mod Z/nZ). Moreover, the element (u0 , v0 ) does not depend on the class of (u, v), i.e., for any s with gcd(N, s) = 1 the input (su, sv) also outputs (u0 , v0 ). If (u, v) is not in P1 (Z/N Z), this algorithm outputs (0, 0), 0. (1) [Reduce] Reduce both u and v modulo N . (2) [Easy (0, 1) Case] If u = 0, check that gcd(v, N ) = 1. If so, return s = 1 and (0, 1); otherwise return 0. (3) [GCD] Compute g = gcd(u, N ) and s, t ∈ Z such that g = su + tN .

146

8. General Modular Symbols

(4) [Not in P 1 ?] We have gcd(u, v, N ) = gcd(g, v), so if gcd(g, v) > 1, then (u, v) 6∈ P1 (Z/N Z), and we return 0. (5) [Pseudo-Inverse] Now g = su+tN , so we may think of s as “pseudoinverse” of u (mod N ), in the sense that su is as close as possible to being 1 modulo N . Note that since g | u, changing s modulo N/g does not change su (mod N ). We can adjust s modulo N/g so it is coprime to N (by adding multiples of N/g to s). (This is because 1 = su/g + tN/g, so s is a unit mod N/g, and the map (Z/N Z)∗ → (Z/(N/g)Z)∗ is surjective, e.g., as we saw in the proof of Algorithm 4.28.) (6) [Multiply by s] Multiply (u, v) by s, and replace (u, v) by the equivalent element (g, sv) of P1 (Z/N Z). (7) [Normalize] Compute the unique pair (g, v ′ ) equivalent to (g, v) that minimizes v, as follows: (a) [Easy Case] If g = 1, this pair is (1, v). (b) [Enumerate and Find Best] Otherwise, note that if 1 6= t ∈ (Z/N Z)∗ and tg ≡ g (mod N ), then (t − 1)g ≡ 0 (mod N ), so t − 1 = kN/g for some k with 1 ≤ k ≤ g − 1. Then for t = 1 + kN/g coprime to N , we have (gt, vt) = (g, v + kvN/g). So we compute all pairs (g, v + kvN/g) and pick out the one that minimizes the least nonnegative residue of vt modulo N . (c) [Invert s and Output] The s that we computed in the above steps multiplies the input (u, v) to give the output (u0 , v0 ). Thus we invert it, since the scalar we output multiplies (u0 , v0 ) to give (u, v). Remark 8.30. In the above algorithm, there are many gcd’s with N so one should create a table of the gcd’s of 0, 1, 2, . . . , N − 1 with N . Remark 8.31. Another approach is to instead use that Y P1 (Z/N Z) ∼ P1 (Z/pνp Z), = p|N

where νp = ordp (N ), and that it is relatively easy to enumerate the elements of P1 (Z/pn Z) for a prime power pn . Algorithm 8.32 (List P1 (Z/N Z)). Given an integer N > 1, this algorithm makes a sorted list of the distinct representatives (c, d) of P1 (Z/N Z) with c 6= 0, 1, as output by Algorithm 8.29. (1) For each c = 1, . . . , N − 1 with g = gcd(c, N ) > 1 do the following: (a) Use Algorithm 8.29 to compute the canonical representative (u′ , v ′ ) equivalent to (c, 1), and include it in the list.

147

8.8. Explicit Examples

(b) If g = c, for each d = 2, . . . , N − 1 with gcd(d, N ) > 1 and gcd(c, d) = 1, append the normalized representative of (c, d) to the list. (2) Sort the list. (3) Pass through the sorted list and delete any duplicates.

8.8. Explicit Examples Explicit detailed examples are crucial when implementing modular symbols algorithms from scratch. This section contains a number of such examples. 8.8.1. Examples of Computation of Mk (Γ0 (N )). In this section, we compute Mk (Γ0 (N )) explicitly in a few cases. Example 8.33. We compute V = M4 (Γ0 (1)). Because Sk (Γ0 (1)) = 0 and Mk (Γ0 (1)) = CE4 , we expect V to have dimension 1 and for each integer n the Hecke operator Tn to have eigenvalue the sum σ3 (n) of the cubes of positive divisors of n. The Manin symbols are x0 = (0, 0, 0), The relation matrix is

x1 = (1, 0, 0),

x2 = (2, 0, 0).



 1 0 1 0 0 0    2 − 2 2 ,    1 − 1 1 2 −2 2 where the first two rows correspond to S-relations and the second three to T -relations. Note that we do not include all S-relations, since it is obvious that some are redundant, e.g., x + xS = 0 and (xS) + (xS)S = xS + x = 0 are the same since S has order 2. The echelon form of the relation matrix is ¶ µ 1 0 1 , 0 1 0

where we have deleted the zero rows from the bottom. Thus we may replace the above complicated list of relations with the following simpler list of relations: x0 + x2 = 0, x1 = 0 from which we immediately read off that the second generator x1 is 0 and x0 = −x2 . Thus M4 (Γ0 (1)) has dimension 1, with basis the equivalence class of x2 (or of x0 ).

148

8. General Modular Symbols

Next we compute the Hecke operator T2 on M4 (Γ0 (1)). The Heilbronn matrices of determinant 2 from Proposition 8.8 are h0 h1 h2 h3

µ 1 = 0 µ 1 = 1 µ 2 = 0 µ 2 = 0

¶ 0 , 2 ¶ 0 , 2 ¶ 0 , 1 ¶ 1 . 1

To compute T2 , we apply each of these matrices to x0 , then reduce modulo the relations. We have µ 1 x2 0 µ 1 x2 1 µ 2 x2 0 µ 2 x2 0

¶ ¶ µ 1 0 0 2 x , = [X , (0, 0)] 0 2 2 2 ¶ 0 = [X 2 , (0, 0)] = x2 , 2 ¶ 0 = [(2X)2 , (0, 0)] = 4x2 , 1 ¶ 1 = [(2X + 1)2 , (0, 0)] = x0 + 4x1 + 4x2 ∼ 3x2 . 1

Summing we see that T2 (x2 ) ∼ 9x2 in M4 (Γ0 (1)). Notice that 9 = 13 + 23 = σ3 (2). The Heilbronn matrices of determinant 3 from Proposition 8.8 are h0 h2 h4 h6

µ 1 = 0 µ 1 = 2 µ 3 = 0 µ 3 = 0

¶ 0 , 3 ¶ 0 , 3 ¶ 0 , 1 ¶ 2 . 1

¶ 1 0 , h1 = 1 3 µ ¶ 2 1 h3 = , 1 2 µ ¶ 3 1 h5 = , 0 1 µ

149

8.8. Explicit Examples

We have µ 1 x2 0 µ 1 x2 1 µ 1 x2 2 µ 2 x2 2 µ 3 x2 0 µ 3 x2 0 µ 3 x2 0

¶ ¶ µ 1 0 0 2 = x2 , = [X , (0, 0)] 0 3 3 ¶ 0 = [X 2 , (0, 0)] = x2 , 3 ¶ 0 = [X 2 , (0, 0)] = x2 , 3 ¶ 1 = [(2X + 1)2 , (0, 0)] = x0 + 4x1 + 4x2 ∼ 3x2 , 2 ¶ 0 = [(3X)2 , (0, 0)] = 9x2 , 1 ¶ 1 = [(3X + 1)2 , (0, 0)] = x0 + 6x1 + 9x2 ∼ 8x2 , 1 ¶ 2 = [(3X + 2)2 , (0, 0)] = 4x0 + 12x1 + 9x2 ∼ 5x2 . 1

Summing we see that T3 (x2 ) ∼ x2 + x2 + x2 + 3x2 + 9x2 + 8x2 + 5x2 = 28x2 . Notice that 28 = 13 + 33 = σ3 (3). Example 8.34. Next we compute M2 (Γ0 (11)) explicitly. The Manin symbol generators are x0 = (0, 1), x1 = (1, 0), x2 = (1, 1), x3 = (1, 2), x4 = (1, 3), x5 = (1, 4), x6 = (1, 5), x7 = (1, 6), x8 = (1, 7), x9 = (1, 8), x10 = (1, 9), x11 = (1, 10). The relation matrix is as follows, and the T -relations are below it:  1 1 0 0 0  0 0 1 0 0   0 0 0 1 0   0 0 0 0 1   0 0 0 0 0   0 0 0 0 0   1 1 0 0 0   0 0 1 0 0   0 0 0 1 0 0 0 0 0 1

where the S-relations are above the line 0 0 0 0 1 0 0 0 1 0

0 0 1 0 0 0 0 1 0 0

0 0 0 0 0 1 0 0 0 1

0 0 0 1 0 0 0 0 1 0

0 0 0 0 1 0 0 0 0 1

0 0 0 0 0 1 0 1 0 0

0 1 0 0 0 0 1 0 0 0



       .       

150

8. General Modular Symbols

In weight 2, two out of three T -relations are redundant, so we do not include them. The reduced row echelon form of the relation matrix is   1 1 0 0 0 0 0 0 0 0 0 0  0 0 1 0 0 0 0 0 0 0 0 0     0 0 0 1 0 0 0 0 0 0 −1 0     0 0 0 0 1 0 0 0 0 1 −1 0     0 0 0 0 0 1 0 0 0 1 0 0  .   0 0 0 0 0 0 1 0 0 0 1 0      0 0 0 0 0 0 0 1 0 0 1 0    0 0 0 0 0 0 0 0 1 −1 1 0     0 0 0 0 0 0 0 0 0 0 0 1  0 0 0 0 0 0 0 0 0 0 0 0 From the echelon form we see that every symbol is equivalent to a combination of x1 = (1, 0), x9 = (1, 8), and x10 = (1, 9). (Notice that columns 1, 9, and 10 are the pivot columns, where we index columns starting at 0.)

To compute T2 , we apply each of the Heilbronn matrices of determinant 2 from Proposition 8.8 to x1 , then to x9 , and finally to x10 . The matrices are as in Example 8.33 above. We have T2 (x1 ) = 3(1, 0) + (1, 6) ∼ 3x1 − x10 . Applying T2 to x9 = (1, 8), we get T2 (x9 ) = (1, 3) + (1, 4) + (1, 5) + (1, 10) ∼ −2x9 . Applying T2 to x10 = (1, 9), we get T2 (x10 ) = (1, 4) + (1, 5) + (1, 7) + (1, 10) ∼ −x1 − 2x10 . Thus the matrix of T2 with respect to this  3 0 T2 =  0 −2 −1 0

basis is  0 0 , −2

where we write the matrix as an operator on the left on vectors written in terms of x1 , x9 , and x10 . The matrix T2 has characteristic polynomial (x − 3)(x + 2)2 . The (x − 3) factor corresponds to the weight 2 Eisenstein series, and the x + 2 factor corresponds to the elliptic curve E = X0 (11), which has a2 = −2 = 2 + 1 − #E(F2 ). Example 8.35. In this example, we compute M6 (Γ0 (3)), which illustrates both weight greater than 2 and level greater than 1. We have the following

151

8.8. Explicit Examples

generating Manin symbols:

x0 = [XY 4 , (0, 1)],

x1 = [XY 4 , (1, 0)],

x2 = [XY 4 , (1, 1)],

x3 = [XY 4 , (1, 2)],

x4 = [XY 3 , (0, 1)],

x5 = [XY 3 , (1, 0)],

x6 = [XY 3 , (1, 1)],

x7 = [XY 3 , (1, 2)],

x8 = [X 2 Y 2 , (0, 1)],

x9 = [X 2 Y 2 , (1, 0)],

x10 = [X 2 Y 2 , (1, 1)],

x11 = [X 2 Y 2 , (1, 2)],

x12 = [X 3 Y, (0, 1)],

x13 = [X 3 Y, (1, 0)],

x14 = [X 3 Y, (1, 1)],

x15 = [X 3 Y, (1, 2)],

x16 = [X 4 Y, (0, 1)],

x17 = [X 4 Y, (1, 0)],

x18 = [X 4 Y, (1, 1)],

x19 = [X 4 Y, (1, 2)].

The relation matrix is already very large for M6 (Γ0 (3)). It is as follows, where the S-relations are before the line and the T -relations after it:

0 B B B B B B B B B B B B B B B B B B B B B B B B B B B B B B B B B B B B B B B B B B B B B B B B B B B @

1 0 0 0 0 0 0 0 0 0 1 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 1

0 1 0 0 0 0 0 0 0 0 0 1 0 1 0 0 0 1 0 0 0 1 0 0 0 1 1 0 0 1

0 0 1 0 0 0 0 0 0 0 0 0 2 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 2 0

0 0 0 1 0 0 0 0 0 0 1 0 0 1 1 0 0 0 1 0 0 0 1 0 0 0 1 1 0 0

0 0 0 0 1 0 0 0 0 0 0 −4 0 0 1 −3 0 0 0 −2 0 0 0 −1 0 −1 0 0 0 −4

0 0 0 0 0 1 0 0 0 0 0 0 0 −4 0 1 0 −3 0 0 0 −2 −1 0 0 −1 −4 0 0 0

0 0 0 0 0 0 1 0 0 0 0 0 −4 0 0 0 −2 0 0 0 −2 0 0 0 −2 0 0 0 −4 0

0 0 0 0 0 0 0 1 0 0 −4 0 0 0 −3 0 0 1 −2 0 0 0 −1 −1 0 0 0 −4 0 0

0 0 0 0 0 0 0 0 1 0 0 6 0 0 0 3 0 0 1 1 0 1 0 0 0 3 0 0 0 6

0 0 0 0 0 0 0 0 1 0 0 0 0 6 0 0 0 3 1 1 0 1 3 0 0 0 6 0 0 0

0 0 0 0 0 0 0 0 0 1 0 0 6 0 0 0 3 0 0 0 3 0 0 0 3 0 0 0 6 0

0 0 0 0 0 0 0 0 0 1 6 0 0 0 3 0 0 0 1 1 0 1 0 3 0 0 0 6 0 0

0 0 0 0 0 −1 0 0 0 0 0 −4 0 0 0 −1 0 −1 0 0 0 −2 1 0 0 −3 0 0 0 −4

0 0 0 0 −1 0 0 0 0 0 0 0 0 −4 −1 0 0 −1 −2 0 0 0 −3 1 0 0 −4 0 0 0

0 0 0 0 0 0 0 −1 0 0 0 0 −4 0 0 0 −2 0 0 0 −2 0 0 0 −2 0 0 0 −4 0

0 0 0 0 0 0 −1 0 0 0 −4 0 0 0 −1 −1 0 0 0 −2 0 0 0 −3 0 1 0 −4 0 0

0 1 0 0 0 0 0 0 0 0 0 1 0 1 0 0 0 1 0 0 0 1 0 0 0 1 1 0 0 1

1 0 0 0 0 0 0 0 0 0 1 0 0 1 1 0 0 0 1 0 0 0 1 0 0 0 1 1 0 0

0 0 0 1 0 0 0 0 0 0 0 0 2 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 2 0

0 0 1 0 0 0 0 0 0 0 1 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 1

1

C C C C C C C C C C C C C C C C C C C C C C C C C C C. C C C C C C C C C C C C C C C C C C C C C C C C A

152

8. General Modular Symbols

The reduced row echelon form of the relations matrix, with zero rows removed is 0 B B B B B B B B B B B B B B B B B B B B B B B B @

1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0

0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0

0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0

0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0

0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0

0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0

0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0

0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0

0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0

0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1

0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0

1 0 0 0 0 0 1/2 −1/2 −1/6 1/6 0 0 0 0 −1/2 1/2

0 0 0 1 3/16 −1/16 −5/16 3/16 1/12 −1/12 1/4 −1/4 −1/16 3/16 3/16 −5/16

0 0 1 0 −3/16 1/16 −3/16 5/16 1/12 −1/12 −1/4 1/4 1/16 −3/16 5/16 −3/16

1

C C C C C C C C C C C C C. C C C C C C C C C C C A

Since these relations are equivalent to the original relations, we see how x0 , . . . , x15 can be expressed in terms of x16 , x17 , x18 , and x19 . Thus M6 (Γ0 (3)) has dimension 4. For example, 1 5 3 x15 ∼ x17 − x18 − x19 . 2 16 16 Notice that the number of relations is already quite large. It is perhaps surprising how complicated the presentation is already for M6 (Γ0 (3)). Because there are denominators in the relations, the above calculation is only a computation of M6 (Γ0 (3); Q). Computing M6 (Γ0 (3); Z) involves finding a Z-basis for the kernel of the relation matrix (see Exercise 7.5). As before, we find that with respect to the basis x16 , x17 , x18 , and x19   33 0 0 0  3 6 12 12   T2 =   −3/2 27/2 15/2 27/2  . −3/2 27/2 27/2 15/2

Notice that there are denominators in the matrix for T2 with respect to this basis. It is clear from the definition of T2 acting on Manin symbols that T2 preserves the Z-module M6 (Γ0 (3)), so there is some basis for M6 (Γ0 (3)) such that T2 is given by an integer matrix. Thus the characteristic polynomial f2 of T2 will have integer coefficients; indeed, f2 = (x − 33)2 · (x + 6)2 . Note the factor (x−33)2 , which comes from the two images of the Eisenstein series E4 of level 1. The factor x + 6 comes from the cusp form g = q − 6q 2 + · · · ∈ S6 (Γ0 (3)). By computing more Hecke operators Tn , we can find more coefficients of g. For example, the charpoly of T3 is (x − 1)(x − 243)(x − 9)2 , and the matrix

153

8.8. Explicit Examples

of T5 is



 3126 0 0 0  240 966 960 960   T5 =   −120 1080 1086 1080  , −120 1080 1080 1086 which has characteristic polynomial f5 = (x − 3126)2 (x − 6)2 .

The matrix of T7 is

 16808 0 0 0  1296 5144 5184 5184   T7 =   −648 5832 5792 5832  , −648 5832 5832 5792 

with characteristic polynomial

f7 = (x − 16808)2 (x + 40)2 .

One can put this information together to deduce that g = q − 6q 2 + 9q 3 + 4q 4 + 6q 5 − 54q 6 − 40q 7 + · · · . Example 8.36. Consider M2 (Γ0 (43)), which has dimension 7. With respect to the symbols x1 = (1, 0),

x32 = (1, 31),

x39 = (1, 38), the matrix of T2 is



    T2 =     

x40 = (1, 39),

x33 = (1, 32), x41 = (1, 40),

x42 = (1, 41),

3 0 0 0 0 0 0 0 −2 −1 −1 −1 0 0 0 1 1 0 0 −2 −1 0 0 1 −1 1 0 0 0 0 0 2 1 2 1 0 0 −1 −1 −1 −2 0 −1 0 0 1 1 1 −1

which has characteristic polynomial



    ,    

(x − 3)(x + 2)2 (x2 − 2)2 .

There is one√Eisenstein series and there are three cusp forms with a2 = −2 and a2 = ± 2. Example 8.37. To compute M2 (Γ0 (2004); Q), we first make a list of the 4032 = (22 + 2) · (3 + 1) · (167 + 1)

elements (a, b) ∈ P1 (Z/2004Z) using Algorithm 8.29. The list looks like this: (0, 1), (1, 0), (1, 1), (1, 2), . . . , (668, 1), (668, 3), (668, 5), (1002, 1).

154

8. General Modular Symbols

For each of the symbols xi , we consider the S-relations and T -relations. Ignoring the redundant relations, we find 2016 S-relations and 1344 T relations. It is simple to quotient out by the S-relations, e.g., by identifying xi with −xi S = −xj for some j (or setting xi = 0 if xi S = xi ). Once we have taken the quotient by the S-relations, we take the image of all 1344 of the T -relations modulo the S-relations and quotient out by those relations. Because S and T do not commutate, we cannot only quotient out by T -relations xi + xi T + xi T 2 = 0 where the xi are the basis after quotienting out by the S-relations. The relation matrix has rank 3359, so M2 (Γ0 (2004); Q) has dimension 673. If we instead compute the quotient M2 (Γ0 (2004); Q)+ of M2 (Γ0 (2004); Q) ∗ by the subspace ¢ elements x − η (x), we include relations xi + xi I = 0, ¡ −1 0 of where I = 0 1 . There are now 2016 S-relations, 2024 I-relations, and 1344 T -relations. Again, it is relatively easy to quotient out by the Srelations by identifying xi and −xi S. We then take the image of all 2024 I-relations modulo the S-relations, and again directly quotient out by the I-relations by identifying [xi ] with −[xi I] = −[xj ] for some j, where by [xi ] we mean the class of xi modulo the S-relations. Finally, we quotient out by the 1344 T -relations, which involves sparse Gauss elimination on a matrix with 1344 rows and at most three nonzero entries per row. The dimension of M2 (Γ0 (2004); Q)+ is 331.

8.9. Refined Algorithm for the Presentation Algorithm 8.38 (Modular Symbols Presentation). This is an algorithm to compute Mk (Γ0 (N ); Q) or Mk (Γ0 (N ); Q)± , which only requires doing generic sparse linear algebra to deal with the three term T -relations. (1) Let x0 , . . . , xn by a list of all Manin symbols. (2) Quotient out the two-term S-relations and if the ± quotient is desired, by the two-term η-relations. (Note that this is more subtle than just “identifying symbols in pairs”, since complicated relations can cause generators to surprisingly equal 0.) Let [xi ] denote the class of xi after this quotienting process. (3) Create a sparse matrix A with m columns, whose rows encode the relations [xi ] + [xi T ] + [xi T 2 ] = 0. For example, there are about n/3 such rows when k = 2. The number of nonzero entries per row is at most 3(k − 1). Note that we must include rows for all i, since even if [xi ] = [xj ], it need not be the case that [xi T ] = [xj T ], since the matrices S and T do not commute. However, we have an a priori formula for the dimension of the quotient by all these relations, so we could omit

8.10. Applications

155

many relations and just check that there are enough at the end—if there are not, we add in more. (4) Compute the reduced row echelon form of A using Algorithm 7.6. For k = 2, this is the echelon form of a matrix with size about n/3 × n/4. (5) Use what we have done above to read off a sparse matrix R that expresses each of the n Manin symbols in terms of a basis of Manin symbols, modulo the relations.

8.10. Applications 8.10.1. Later in This Book. We sketch some of the ways in which we will apply the modular symbols algorithms of this chapter later in this book. Cuspidal modular symbols are in Hecke-equivariant duality with cuspidal modular forms, and as such we can compute modular forms by computing systems of eigenvalues for the Hecke operators acting on modular symbols. By the Atkin-Lehner-Li theory of newforms (see, e.g., Theorem 9.4), we can construct Sk (N, ε) for any N , any ε, and k ≥ 2 using this method. See Chapter 1 for more details. Once we can compute spaces of modular symbols, we move to computing the corresponding modular forms. We define inclusion and trace maps from modular symbols of one level N to modular symbols of level a multiple or divisor of N . Using these, we compute the quotient V of the new subspace of cuspidal modular symbols on which a “star involution” acts as +1. The Hecke operators act by diagonalizable commuting matrices on this space, and computing systems of Hecke eigenvalues is equivalent to computing P the n newforms an q . In this way, we obtain a list of all newforms (normalized eigenforms) in Sk (N, ε) for any N , ε, and k ≥ 2.

In Chapter 10, we compute with the period mapping from modular symbols to C attached to a newform f ∈ Sk (N, ε). When k = 2, ε = 1 and f has rational Fourier coefficients, this gives a method to compute the period lattice associated to a modular elliptic curve attached to a newform (see Section 10.7). In general, computation of this map is important when finding equations for modular Q-curves, CM curves, and curves with a given modular Jacobian. It is also important for computing special values of the L-function L(f, s) at integer points in the critical strip.

8.10.2. Discussion of the Literature and Research. Modular symbols were introduced by Birch [Bir71] for computations in support of the Birch and Swinnerton-Dyer conjecture. Manin [Man72] used modular symbols to prove rationality results about special values of L-functions.

156

8. General Modular Symbols

Merel’s paper [Mer94] builds on work of Shokurov (mainly [Sho80a]), which develops a higher-weight generalization of Manin’s work partly to understand rationality properties of special values of L-functions. Cremona’s book [Cre97a] discusses how to compute the space of weight 2 modular symbols for Γ0 (N ), in connection with the problem of enumerating all elliptic curves of given conductor, and his article [Cre92] discusses the Γ1 (N ) case and computation of modular symbols with character. There have been several Ph.D. theses about modular symbols. Basmaji’s thesis [Bas96] contains tricks to efficiently compute Hecke operators Tp , with p very large (see Section 8.3.4), and also discusses how to compute spaces of half integral weight modular forms building on what one can get from modular symbols of integral weight. The author’s Ph.D. thesis [Ste00] discusses higher-weight modular symbols and applies modular symbols to study Shafarevich-Tate groups (see also [Aga00]). Martin’s thesis [Mar01] is about an attempt to study an analogue of analytic modular symbols for weight 1. Gabor Wiese’s thesis [Wie05] uses modular symbols methods to study weight 1 modular forms modulo p. Lemelin’s thesis [Lem01] discusses modular symbols for quadratic imaginary fields in the context of p-adic analogues of the Birch and Swinnerton-Dyer conjecture. See also the survey paper [FM99], which discusses computation with weight 2 modular symbols in the context of modular abelian varieties. The appendix of this book is about analogues of modular symbols for groups besides finite index subgroups of SL2 (Z), e.g., for subgroup of higher rank groups such as SL3 (Z). There has also been work on computing Hilbert modular forms, e.g., by Lassina Dembel´e [Dem05] Hilbert modular forms are functions on a product of copies of h, and SL2 (Z) is replaced by a group of matrices with entries in a totally real field. Glenn Stevens, Robert Pollack and Henri Darmon (see [DP04]) have worked for many years to develop an analogue of modular symbols in a rigid analytic context, which is helpful for questions about computing with overconvergent p-adic modular forms or proving results about p-adic L-functions. Finally we mention that Barry Mazur and some other authors use the term “modular symbol” in a different way than we do. They use the term in a way that is dual to our usage; for example, they attach a “modular symbol” to a modular form or elliptic curve. See [MTT86] for an extensive discussion of modular symbols from this point of view, where they are used to construct p-adic L-functions.

8.11. Exercises 8.1 Suppose M is an integer multiple of N . Prove that the natural map (Z/M Z)∗ → (Z/N Z)∗ is surjective.

8.11. Exercises

157

8.2 Prove that SL2 (Z) → SL2 (Z/N Z) is surjective (see Lemma 8.15).

8.3 Compute M3 (Γ1 (3)). List each Manin symbol the relations they satisfy, compute the quotient, etc. Find the matrix of T2 . (Check: The dimension of M3 (Γ1 (3)) is 2, and the characteristic polynomial of T2 is (x − 3)(x + 3).)

8.4 Finish the proof of Proposition 8.17. ¡ 0¢ 8.5 (a) Show that if η = −1 0 1 , then ηΓη = Γ for Γ = Γ0 (N ) and Γ = Γ1 (N ). (b) (*) Give an example of a finite index subgroup Γ such that ηΓη 6= Γ.

Chapter 9

Computing with Newforms

In this chapter we pull together results and algorithms from Chapter 3, 4, 7, and 8 and explain how to use linear algebra techniques to compute cusp forms and eigenforms using modular symbols. We first discuss in Section 9.1 how to decompose Mk (Γ1 (N )) as a direct sum of subspaces corresponding to Dirichlet characters. Next in Section 9.2 we state the main theorems of Atkin-Lehner-Li theory, which decomposes Sk (Γ1 (N )) into subspaces on which the Hecke operators act diagonalizably with “multiplicity one”. In Section 9.3 we describe two algorithms for computing modular forms. One algorithm finds a basis of q-expansions, and the other computes eigenvalues of newforms.

9.1. Dirichlet Character Decomposition The group (Z/N Z)∗ acts on Mk (Γ1 (N )) through diamond-bracket operators hdi, as follows. For d ∈ (Z/N Z)∗ , define a b c d′

”i

k

,

¢ −1 where c db′ ∈ SL2 (Z) is congruent to d0 d0 (mod N ). Note that the ¢SL2 (Z) → SL2 (Z/N Z) is surjective (see Exercise 8.2), so the matrix ¡map a b exists. To prove that hdi preserves Mk (Γ1 (N )), we prove the more ′ c d general fact that Γ1 (N ) is a normal subgroup of ¶ ¾ ¶ µ ¶ µ ½µ ∗ ∗ a b a b (mod N ) . ≡ ∈ SL2 (Z) : Γ0 (N ) = 0 ∗ c d c d ¡a

¢

f |hdi = f

h“

¡

159

160

9. Computing with Newforms

This will imply that hdi preserves Mk (Γ1 (N )) since

¡a

b c d′

¢

∈ Γ0 (N ).

Lemma 9.1. The group Γ1 (N ) is a normal subgroup of Γ0 (N ), and the quotient Γ0 (N )/Γ1 (N ) is isomorphic to (Z/N Z)∗ . Proof. See Exercise 9.1.

¤

Alternatively, one can prove that hdi preserves Mk (Γ1 (N )) by showing that hdi ∈ T and noting that Mk (Γ1 (N )) is preserved by T (see Remark 9.11). The diamond-bracket action is the action of Γ0 (N )/Γ1 (N ) ∼ = (Z/N Z)∗

on Mk (Γ1 (N )). Since Mk (Γ1 (N )) is a finite-dimensional vector space over C, the hdi action breaks Mk (Γ1 (N )) up as a direct sum of factors corresponding to the Dirichlet characters D(N, C) of modulus N .

Proposition 9.2. We have Mk (Γ1 (N )) =

M

Mk (N, ε),

ε∈D(N,C)

where

© ª Mk (N, ε) = f ∈ Mk (Γ1 (N )) : f |hdi = ε(d)f, all d ∈ (Z/N Z)∗ .

Proof. The linear transformations hdi, for the d ∈ (Z/N Z)∗ , all commute, since hdi acts through the abelian group Γ0 (N )/Γ1 (N ). Also, if e is the exponent of (Z/N Z)∗ , then hdie = hde i = h1i = 1, so the matrix of hdi is diagonalizable. It is a standard fact from linear algebra that any commuting family of diagonalizable linear transformations is simultaneously diagonalizable (see Exercise 5.1), so there is a basis f1 , . . . , fn for Mk (Γ1 (N )) such that all hdi act by diagonal matrices. The system of eigenvalues of the action of (Z/N Z)∗ on a fixed fi defines a Dirichlet character, i.e., each fi has the property that fi |hdi = εi (d), for all d ∈ (Z/N Z)∗ and some Dirichlet character εi . The fi for a given ε then span Mk (N, ε), and taken together the Mk (N, ε) must span Mk (Γ1 (N )). ¤ Definition 9.3 (Character of Modular Form). If f ∈ Mk (N, ε), we say that ε is the character of the modular form f . The spaces Mk (N, ε) are a direct sum of subspaces Sk (N, ε) and Ek (N, ε), where Sk (N, ε) is the subspace of cusp forms, i.e., forms that vanish at all cusps (elements of Q ∪ {∞}), and Ek (N, ε) is the subspace of Eisenstein series, which is the unique subspace of Mk (N, ε) that is invariant under all Hecke operators and is such that Mk (N, ε) = Sk (N, ε)⊕Ek (N, ε). The space Ek (N, ε) can also be defined as the space spanned by all Eisenstein series of weight k and level N , as defined in Chapter 5. The space Ek (N, ε) can

9.2. Atkin-Lehner-Li Theory

161

be defined in a third way using the Petersson inner product (see [Lan95, §VII.5]).

The diamond-bracket operators preserve cusp forms, so the isomorphism of Proposition 9.2 restricts to an isomorphism of the corresponding cuspidal subspaces. We illustrate how to use SAGE to make a table of dimension of Mk (Γ1 (N )) and Mk (N, ε) for N = 13. sage: G = DirichletGroup(13) sage: G Group of Dirichlet characters of modulus 13 over Cyclotomic Field of order 12 and degree 4 sage: dimension_modular_forms(Gamma1(13),2) 13 sage: [dimension_modular_forms(e,2) for e in G] [1, 0, 3, 0, 2, 0, 2, 0, 2, 0, 3, 0]

Next we do the same for N = 100. sage: G = DirichletGroup(100) sage: G Group of Dirichlet characters of modulus 100 over Cyclotomic Field of order 20 and degree 8 sage: dimension_modular_forms(Gamma1(100),2) 370 sage: v = [dimension_modular_forms(e,2) for e in G]; v [24, 0, 0, 17, 18, 0, 0, 17, 18, 0, 0, 21, 18, 0, 0, 17, 18, 0, 0, 17, 24, 0, 0, 17, 18, 0, 0, 17, 18, 0, 0, 21, 18, 0, 0, 17, 18, 0, 0, 17] sage: sum(v) 370

9.2. Atkin-Lehner-Li Theory In Section 8.6 we defined maps between modular symbols spaces of different level. There are similar maps between spaces of cusp forms. Suppose N and M are positive integers with M | N and that t is a divisor of N/M . Let (9.2.1)

αt : Sk (Γ1 (M )) → Sk (Γ1 (N ))

be the degeneracy map, which is given by f (q) 7→ f (q t ). There are also maps βt in the other direction; see [Lan95, Ch. VIII].

162

9. Computing with Newforms

The old subspace of Sk (Γ1 (N )), denoted Sk (Γ1 (N ))old , is the sum of the images of all maps αt with M a proper divisor of N and t any divisor of N/M (note that αt depends on t, N , and M , so there is a slight abuse of notation). The new subspace of Sk (Γ1 (N )), which we denote by Sk (Γ1 (N ))new , is the intersection of the kernel of all maps βt with M a proper divisor of N . One can use the Petersson inner product to show that Sk (Γ1 (N )) = Sk (Γ1 (N ))new ⊕ Sk (Γ1 (N ))old .

Moreover, the new and old subspaces are preserved by all Hecke operators. Let T = Z[T1 , T2 , . . .] be the commutative polynomial ring in infinitely many indeterminates Tn . This ring acts (via Tn acting as the nth Hecke operator) on Sk (Γ1 (N )) for every integer N . Let T(N ) be the subring of T generated by the Tn with gcd(n, N ) = 1. Theorem 9.4 (Atkin, Lehner, Li). We have a decomposition M M αd (Sk (Γ1 (M ))new ). (9.2.2) Sk (Γ1 (N )) = M |N d|N/M

Each space Sk (Γ1 (M ))new is a direct sum of distinct (nonisomorphic) simple (N ) TC -modules. Proof. The complete proof is in [Li75]. See also [DS05, Ch. 5] for a beautiful modern treatment of this and related results. ¤ The analogue of Theorem 9.4 with Γ1 replaced by Γ0 is also true (this is what was proved in [AL70]). The analogue for Sk (N, ε) is also valid, as long as we omit the spaces Sk (Γ1 (M ), ε) for which cond(ε) ∤ M . Example 9.5. If N is prime and k ≤ 11, then Sk (Γ1 (N ))new = Sk (Γ1 (N )), since Sk (Γ1 (1)) = 0. One can prove using the Petersson inner product that the operators Tn on Sk (Γ1 (N )), with gcd(n, N ) = 1, are diagonalizable. Another result of Atkin-Lehner-Li theory is that the ring of endomorphisms of Sk (Γ1 (N ))new generated by all Hecke operators equals the ring generated by the Hecke operators Tn with (n, N ) = 1. This statement need not be true if we do not restrict to the new subspace, as the following example shows. Example 9.6. We have S2 (Γ0 (22)) = S2 (Γ0 (11)) ⊕ α2 (S2 (Γ0 (11))),

where each of the spaces S2 (Γ0 (11)) has dimension 1. Thus S2 (Γ0 (22))new = 0. The Hecke operator T2 on S2 (Γ0 (22)) has characteristic polynomial x2 + 2x + 2, which is irreducible. Since α2 commutes with all Hecke operators Tn , with gcd(n, 2) = 1, the subring T(22) of the Hecke algebra generated by

9.2. Atkin-Lehner-Li Theory

163

operators Tn with n odd is isomorphic to Z (the 2 × 2 scalar matrices). Thus on the full space S2 (Γ0 (22)), we do not have T(22) = T. However, on the new subspace we do have this equality, since the new subspace has dimension 0. Example 9.7. The space S2 (Γ0 (45)) has dimension 3 and basis f0 = q − q 4 − q 10 − 2q 13 − q 16 + 4q 19 + · · · , f1 = q 2 − q 5 − 3q 8 + 4q 11 − 2q 17 + · · · ,

f2 = q 3 − q 6 − q 9 − q 12 + q 15 + q 18 + · · · . The new subspace S2 (Γ0 (45))new is spanned by the single cusp form q + q 2 − q 4 − q 5 − 3q 8 − q 10 + 4q 11 − 2q 13 + · · · . We have S2 (Γ0 (45/5)) = 0 and S2 (Γ0 (15)) has dimension 1 with basis q − q 2 − q 3 − q 4 + q 5 + q 6 + 3q 8 + q 9 − q 10 − 4q 11 + q 12 − 2q 13 + · · · . We use SAGE to verify the above assertions. sage: S = CuspForms(Gamma0(45), 2, prec=14); S Cuspidal subspace of dimension 3 of Modular Forms space of dimension 10 for Congruence Subgroup Gamma0(45) of weight 2 over Rational Field sage: S.basis() [ q - q^4 - q^10 - 2*q^13 + O(q^14), q^2 - q^5 - 3*q^8 + 4*q^11 + O(q^14), q^3 - q^6 - q^9 - q^12 + O(q^14) ] sage: S.new_subspace().basis() (q - q^4 - q^10 - 2*q^13 + O(q^14),) sage: CuspForms(Gamma0(9),2) Cuspidal subspace of dimension 0 of Modular Forms space of dimension 3 for Congruence Subgroup Gamma0(9) of weight 2 over Rational Field sage: CuspForms(Gamma0(15),2, prec=10).basis() [ q - q^2 - q^3 - q^4 + q^5 + q^6 + 3*q^8 + q^9 + O(q^10) ] Example 9.8. This example is similar to Example 9.6, except that there are newforms. We have S2 (Γ0 (55)) = S2 (Γ0 (11)) ⊕ α5 (S2 (Γ0 (11))) ⊕ S2 (Γ0 (55))new ,

164

9. Computing with Newforms

where S2 (Γ0 (11)) has dimension 1 and S2 (Γ0 (55))new has dimension 3. The Hecke operator T5 on S2 (Γ0 (55))new acts via the matrix   −2 2 −1 −1 1 −1 1 −2 0

with respect to some basis. This matrix has eigenvalues 1 and −1. AtkinLehner theory asserts that T5 must be a linear combination of Tn , with gcd(n, 55) = 1. Upon computing the matrix for T2 , we find by simple linear algebra that T5 = 2T2 − T4 . Definition 9.9 (Newform). A newform is a T-eigenform f ∈ Sk (Γ1 (N ))new that is normalized so that the coefficient of q is 1. We now motivate this definition by explaining why any T-eigenform can be normalized so that the coefficient of q is 1 and how such an eigenform has the property that its Fourier coefficients are exactly the Hecke eigenvalues. P n Proposition 9.10. If f = ∞ n=0 an q ∈ Mk (N, ε) is an eigenvector for all Hecke operators Tn normalized so that a1 = 1, then Tn (f ) = an f . Proof. If ε = 1, then f ∈ Mk (Γ0 (N )) and this is Lemma 3.22. However, we have not yet considered Hecke operators on q-expansions for more general spaces of modular forms. The Hecke operators Tp , for p prime, act on Sk (N, ε) by ! Ã∞ ∞ ³ ´ X X n an q anp q n + ε(p)pk−1 an q np , = Tp n=0

n=0

P n and there is a similar formula for Tm with m composite. If f = ∞ n=0 an q is an eigenform for all Tp , with eigenvalues λp , then by the above formula (9.2.3)

λp f = λp a1 q + λp a2 q 2 + · · · = Tp (f ) = ap q + higher terms.

Equating coefficients of q, we see that if a1 = 0, then ap = 0 for all p; hence an = 0 for all n, because of the multiplicativity of Fourier coefficients and the recurrence apr = apr−1 ap − ε(p)pk−1 apr−2 . This would mean that f = 0, a contradiction. Thus a1 6= 0, and it makes sense to normalize f so that a1 = 1. With this normalization, (9.2.3) implies that λp = ap , as desired. ¤ Remark 9.11. The Hecke algebra TQ on Mk (Γ1 (N )) contains the operators hdi, since they satisfy the relation Tp2 = Tp2 −hpipk−1 . Thus any T-eigenform in Mk (Γ1 (N )) lies in a subspace Mk (N, ε) for some Dirichlet character ε. Also, one can even prove that hdi ∈ Z[. . . , Tn , . . .] (see Exercise 9.2).

9.3. Computing Cusp Forms

165

9.3. Computing Cusp Forms Let Sk (N, ε; C) be the space of cuspidal modular symbols as in Chapter 8. Let ι∗ be the map of (8.5.8), and let Sk (N, ε; C)+ be the plus one quotient of cuspidal modular symbols, i.e., the quotient of Sk (N, ε; C) by the image of ι∗ − 1. It follows from Theorem 8.23 and compatibility of the degeneracy maps (for modular symbols they are defined in Section 8.6) that the Tmodules Sk (N, ε)new and Sk (N, ε; C)+ new are dual as T-modules. Thus finding the systems of T-eigenvalues on cusp forms is the same as finding the systems of T-eigenvalues on cuspidal modular symbols. Our strategy to compute Sk (N, ε) is to first compute spaces Sk (N, ε)new using the Atkin-Lehner-Li decomposition (9.2.2). To compute Sk (N, ε)new to a given precision, we compute the systems of eigenvalues of the Hecke operators Tp on the space V = Sk (N, ε; C)+ new , which we will define below. Using Proposition 9.10, we then recover a basis of q-expansions for newforms. Note that we only need to compute Hecke eigenvalues Tp , for p prime, not the Tn for n composite, since the an can be quickly recovered in terms of the ap using multiplicativity and the recurrence. For some problems, e.g., construction of models for modular curves, having a basis of q-expansions is enough. For many other problems, e.g., enumeration of modular abelian varieties, one is really interested in the newforms. We next discuss algorithms aimed at each of these problems. 9.3.1. A Basis of q-Expansions. The following algorithm generalizes Algorithm 3.26. It computes Sk (N, ε) without finding any eigenspaces. Algorithm 9.12 (Merel’s Algorithm for Computing a Basis). Given integers m, N and k and a Dirichlet character ε with modulus N , this algorithm computes a basis of q-expansions for Sk (N, ε) to precision O(q m+1 ). (1) [Compute Modular Symbols] Use Algorithm 8.38 to compute V = Sk (N, ε)+ ⊗ Q(ε), viewed as a K = Q(ε) vector space, with an action of the Tn . (2) [Basis for Linear Dual] Write down a basis for V ∗ = Hom(V, Q(ε)). E.g., if we identify V with K n viewed as column vectors, then V ∗ is the space of row vectors of length n, and the pairing is the row × column product. (3) [Find Generator] Find x ∈ V such that Tx = V by choosing random x until we find one that generates. The set of x that fail to generate lie in a union of a finite number of proper subspaces.

166

9. Computing with Newforms

(4) [Compute Basis] The set of power series fi =

m X

ψi (Tn (x))q n + O(q m+1 )

n=1

forms a basis for Sk (N, ε) to precision m. In practice Algorithm 9.12 seems slower than the eigenspace algorithm that we will describe in the rest of this chapter. The theoretical complexity of Algorithm 9.12 may be better, because it is not necessary to factor any polynomials. Polynomial factorization is difficult from the worst-case complexity point of view, though it is usually fast in practice. The eigenvalue algorithm only requires computing a few images Tp (x) for p prime and x a Manin symbol on which Tp can easily be computed. The Merel algorithm involves computing Tn (x) for all n and for a fairly easy x, which is potentially more work. Remark 9.13. By “easy x”, I mean that computing Tn (x) is easier on x than on a completely random element of Sk (N, ε)+ , e.g., x could be a Manin symbol. 9.3.2. Newforms: Systems of Eigenvalues. In this section we describe an algorithm for computing the system of Hecke eigenvalues associated to a simple subspace of a space of modular symbols. This algorithm is better than doing linear algebra directly over the number field generated by the eigenvalues. It only involves linear algebra over the base field and also yields a compact representation for the answer, which is better than writing the eigenvalues in terms of a power basis for a number field. In order to use this algorithm, it is necessary to decompose the space of cuspidal modular symbols as a direct sum of simples, e.g., using Algorithm 7.17. Fix N and a Dirichlet character ε of modulus N , and let V = Mk (N, ε)+ be the +1 quotient of modular symbols (see equation (8.5.8)). Algorithm 9.14 (System of Eigenvalues). Given a T-simple subspace W ⊂ V of modular symbols, this algorithm outputs maps ψ and e, where ψ : TK → W is a K-linear map and e : W ∼ = L is an isomorphism of W with a number field L, such that an = e(ψ(Tn )) is the eigenvalue of theP nth Hecke operator n acting on a fixed T-eigenvector in W ⊗ Q. (Thus f = ∞ n=1 e(ψ(Tn ))q is a newform.)

(1) [Compute Projection] Let ϕ : V → W ′ be any surjective linear map such that ker(ϕ) equals the kernel of the T-invariant projection onto W . For example, compute ϕ by finding a simple submodule

167

9.3. Computing Cusp Forms

of V ∗ = Hom(V, K) that is isomorphic to W , e.g., by applying Algorithm 7.17 to V ∗ with T replaced by the transpose of T . (2) [Choose v] Choose a nonzero element v ∈ V such that π(v) 6= 0 and computation of Tn (v) is “easy”, e.g., choose v to be a Manin symbol. (3) [Map from Hecke Ring] Let ψ be the map T → W ′ , given by ψ(t) = π(tv). Note that computation of ψ is relatively easy, because v was chosen so that tv is relatively easy to compute. In particular, if t = Tp , we do not need to compute the full matrix of Tp on V ; instead we just compute Tp (v). (4) [Find Generator] Find a random T ∈ T such that the iterates ψ(T 0 ),

ψ(T ),

ψ(T 2 ),

...,

ψ(T d−1 )

are a basis for W ′ , where W has dimension d. (5) [Characteristic Polynomial] Compute the characteristic polynomial f of T |W , and let L = K[x]/(f ). Because of how we chose T in step (4), the minimal and characteristic polynomials of T |W are equal, and both are irreducible, so L is an extension of K of degree d = dim(W ). (6) [Field Structure] In this step we endow W ′ with a field structure. Let e : W ′ → L be the unique K-linear isomorphism such that e(ψ(T i )) ≡ xi

(mod f )

for i = 0, 1, 2, ..., deg(f ) − 1. The map e is uniquely determined since the ψ(T i ) are a basis for W ′ . To compute e, we compute the change of basis matrix from the standard basis for W ′ to the basis {ψ(T i )}. This change of basis matrix is the inverse of the matrix whose rows are the ψ(T i ) for i = 0, ..., deg(f ) − 1. (7) [Hecke Eigenvalues] Finally for each integer n ≥ 1, we have an = e(ψ(Tn )) = e(π(Tn (v))), where an is the eigenvalue of Tn . Output the maps ψ and e and terminate. One reason we separate ψ and e is that when dim(W ) is large, the values ψ(Tn ) take less space to store and are easier to compute, whereas each one of the values e(ψ(n)) is huge.1 The function e typically involves large numbers if dim(W ) is large, since e is obtained from the iterates of a single vector. For many applications, e.g., databases, it is better to store a matrix that defines e and the images under ψ of many Tn . 1John Cremona initially suggested to me the idea of separating these two maps.

168

9. Computing with Newforms

Example 9.15. The space S2 (Γ0 (23)) of cusp forms has dimension 2 and is spanned by two Gal(Q/Q)-conjugate newforms, one of which is f = q + aq 2 + (−2a − 1)q 3 + (−a − 1)q 4 + 2aq 5 + · · · , √ where a = (−1 + 5)/2. We will use Algorithm 9.14 to compute a few of these coefficients. The space M2 (Γ0 (23))+ of modular symbols has dimension 3. It has the following basis of Manin symbols: [(0, 0)],

[(1, 0)],

[(0, 1)],

where we use square brackets to differentiate Manin symbols from vectors. The Hecke operator   3 0 0 0 2  T2 =  0 −1 1/2 −1

has characteristic polynomial (x − 3)(x2 + x − 1). The kernel of T2 − 3 corresponds to the span of the Eisenstein series of level 23 and weight 2, and the kernel V of T22 + T2 − 1 corresponds to S2 (Γ0 (23)). (We could also have computed V as the kernel of the boundary map M2 (Γ0 (23))+ → B2 (Γ0 (23))+ .) Each of the following steps corresponds to the step of Algorithm 9.14 with the same number. (1) [Compute Projection] We compute projection onto V (this will suffice to give us a map φ as in the algorithm). The matrix whose first two columns are the echelon basis for V and whose last column is the echelon basis for the Eisenstein subspace is   0 0 1  1 0 −2/11  0 1 −3/11 and

B −1



 2/11 1 0 =  3/11 0 1  , 1 0 0

so projection onto V is given by the first two rows: ¶ µ 2/11 1 0 . π= 3/11 0 1 (2) [Choose v] Let v = (0, 1, 0)t . Notice that π(v) = (1, 0)t 6= 0, and v = [(1, 0)] is a sum of only one Manin symbol.

9.3. Computing Cusp Forms

169

(3) [Map from Hecke Ring] This step is purely conceptual, since no actual work needs to be done. We illustrate it by computing ψ(T1 ) and ψ(T2 ). We have ψ(T1 ) = π(v) = (1, 0)t and ψ(T2 ) = π(T2 (v)) = π((0, 0, 1/2)t ) = (0, 1/2)t . (4) [Find Generator] We have ψ(T20 ) = ψ(T1 ) = (1, 0)t , which is clearly independent from ψ(T2 ) = (0, 1/2)t . Thus we find that the image of the powers of T = T2 generate V . ³ ´ 0 2 (5) [Characteristic Polynomial] The matrix of T2 |V is 1/2 −1 , which

has characteristic polynomial f = x2 + x − 1. Of course, we already knew this because we computed V as the kernel of T22 + T2 − 1.

(6) [Field Structure] We have

ψ(T20 ) = π(v) = (1, 0)t and ψ(T2 ) = (0, 1/2). ³ ´ 0 The matrix with rows the ψ(T2i ) is 10 1/2 , which has inverse e =

( 10 02 ). The matrix e defines an isomorphism between V and the field √ L = Q[x]/(f ) = Q((−1 + 5)/2). √ I.e., e((1, 0)) = 1 and e((0, 1)) = 2x, where x = (−1 + 5)/2. (7) [Hecke Eigenvalues] We have an = e(Ψ(Tn )). For example, a1 = e(Ψ(T1 )) = e((1, 0)) = 1, a2 = e(Ψ(T2 )) = e((0, 1/2)) = x, a3 = e(Ψ(T3 )) = e(π(T3 (v))) = e(π((0, −1, −1)t )) = e((−1, −1)t ) = −1 − 2x,

a4 = e(Ψ(T4 )) = e(π((0, −1, −1/2)t )) = e((−1, −1/2)t ) = −1 − x,

a5 = e(Ψ(T5 )) = e(π((0, 0, 1)t )) = e((0, 1)t ) = 2x, a23 = e(Ψ(T23 )) = e(π((0, 1, 0)t )) = e((1, 0)t ) = 1,

a97 = e(Ψ(T23 )) = e(π((0, 14, 3)t )) = e((14, 3)t ) = 14 + 6x. Example 9.16. It is easier to appreciate Algorithm 9.14 after seeing how big the coefficients of the power series expansion of a newform typically are,

170

9. Computing with Newforms

when the newform is defined over a large field. For example, there is a newform ∞ X an q n ∈ S2 (Γ0 (389)) f= n=1

such that if α = a2 , then

1097385680·a3 (f ) = −20146763x19 + 102331615x18 + 479539092x17 − 3014444212x16 − 3813583550x15 + 36114755350x14

+ 6349339639x13 − 227515736964x12 + 71555185319x11

+ 816654992625x10 − 446376673498x9 − 1698789732650x8

+ 1063778499268x7 + 1996558922610x6 − 1167579836501x5 − 1238356001958x4 + 523532113822x3 + 352838824320x2 − 58584308844x − 25674258672.

In contrast, if we take v = {0, ∞} = (0, 1) ∈ M2 (Γ0 (389))+ , then

T3 (v) = −4(1, 0) + 2(1, 291) − 2(1, 294) − 2(1, 310) + 2(1, 313) + 2(1, 383).

Storing T3 (v), T5 (v), . . . as vectors is more compact than storing a3 (f ), a5 (f ), . . . directly as polynomials in a2 !

9.4. Congruences between Newforms This section is about congruences between modular forms. Understanding congruences is crucial for studying Serre’s conjectures, Galois representations, and explicit construction of Hecke algebras. We assume more background in algebraic number theory here than elsewhere in this book. 9.4.1. Congruences between Modular Forms. Let Γ be an arbitrary congruence subgroup of SL2 (Z), and f ∈ Mk (Γ) is a modular form ¡ suppose ¢ of integer weight k for Γ. Since 10 N1 ∈ Γ for some integer N , the form f has a Fourier expansion in nonnegative powers of q 1/N . For a rational number n, let an (f ) be the coefficient of q n in the Fourier expansion of f . Put ordq (f ) = min{n ∈ Q : an 6= 0}, where by convention we take min ∅ = +∞, so ordq (0) = +∞.

9.4.1.1. The j-invariant. Let 1 j = + 744 + 196884q + · · · q be the j-function, which is a weight 0 modular function that is holomorphic except for a simple pole at ∞ and has integer Fourier coefficients (see, e.g., [Ser73, Section VIII.3.3]).

171

9.4. Congruences between Newforms

Lemma 9.17. Suppose g is a weight 0 level 1 modular function that is holomorphic except possibly with a pole of order n at ∞. Then g is a polynomial in j of degree at most n. Moreover, the coefficients of this polynomial lie in the ideal I generated by the coefficients am (g) with m ≤ 0. Proof. If n = 0, then g ∈ M0 (SL2 (Z)) = C, so g is constant with constant term in I, so the statement is true. Next suppose n > 0 and the lemma has been proved for all functions with smaller order poles. Let α = an (g), and note that µ µ ¶n ¶ 1 n + 744 + 196884q + · · · > −n. ordq (g − αj ) = ordq g − α · q

Thus by induction h = g − αj n is a polynomial in j of degree < n with coefficients in the ideal generated by the coefficients am (g) with m < 0. It follows that g = α · j n − h satisfies the conclusion of the lemma. ¤ 9.4.1.2. Sturm’s Theorem. If O is the of integers of a number field, m P ring is a maximal ideal of O, and f = an q n ∈ O[[q 1/N ]] for some integer N , let ordm(f ) = ordq (f mod m) = min{n ∈ Q : an 6∈ m}.

Note that ordm(f g) = ordm(f ) + ordm(g). The following theorem was first proved in [Stu87]. Theorem 9.18 (Sturm). Let m be a prime ideal in the ring of integers O of a number field K, and let Γ be a congruence subgroup of SL2 (Z) of index m and level N . Suppose f ∈ Mk (Γ, O) is a modular form and ordm(f ) >

or f ∈ Sk (Γ, O) is a cusp form and ordm(f ) >

km 12

km m − 1 − . 12 N

Then f ≡ 0 (mod m). Proof. Case 1: First we assume Γ = SL2 (Z). Let ∆ = q + 24q 2 + · · · ∈ S12 (SL2 (Z), Z)

be the ∆ function. Since ordm(f ) > k/12, we have ordm(f 12 ) > k. We have

(9.4.1)

ordq (f 12 · ∆−k ) = 12 · ordq (f ) − k · ordq (∆) ≥ −k,

since f is holomorphic at infinity and ∆ has a zero of order 1. Also (9.4.2)

ordm(f 12 · ∆−k ) = ordm(f 12 ) − k · ordm(∆) > k − k = 0.

172

9. Computing with Newforms

Combining (9.4.1) and (9.4.2), we see that X bn q n , f 12 · ∆−k = n≥−k

with bn ∈ O and bn ∈ m if n ≤ 0. By Lemma 9.17,

f 12 · ∆−k ∈ m[j] is a polynomial in j of degree at most k with coefficients in m. Thus f 12 ∈ m[j] · ∆k ,

so since the coefficients of ∆ are integers, every coefficient of f 12 is in m. Thus ordm(f 12 ) = +∞, hence ordm(f ) = +∞, so f = 0, as claimed. Case 2: Γ Arbitrary Let N be such that Γ(N ) ⊂ Γ, so also f ∈ Mk (Γ(N )). If g ∈ Mk (Γ(N )) is arbitrary, then because Γ(N ) is a normal subgroup of SL2 (Z), we have that for any γ ∈ Γ(N ) and δ ∈ SL2 (Z), ′

′

(g [δ]k )[γ]k = g [δγ]k = g [γ δ]k = (g [γ ]k )[δ]k = g [δ]k ,

where γ ′ ∈ SL2 (Z). Thus for any δ ∈ SL2 (Z), we have that g [δ]k ∈ Mk (Γ(N )), so SL2 (Z) acts on Mk (Γ(N )). It is a standard (but nontrivial) fact about modular forms, which comes from the geometry of the modular curve X(N ) over Q(ζN ) and Z[ζN ], that Mk (Γ(N )) has a basis with Fourier expansions in Z[ζN ][[q 1/N ]] and that the action of SL2 (Z) on Mk (Γ(N )) preserves Mk (Γ(N ), Q(ζN )) = Mk (Γ(N )) ∩ (Q(ζN )[[q 1/N ]])

and the cuspidal subspace Sk (Γ(N ), Q(ζN )). In particular, for any γ ∈ SL2 (Z), f [γ]k ∈ Mk (Γ(N ), K(ζN ))

Moreover, the denominators of f [γ]k are bounded, since f is an O[ζN ]-linear combination of a basis for Mk (Γ(N ), Z[ζN ]), and the denominators of f [γ]k divide the product of the denominators of the images of each of these basis vectors under [γ]k . Let L = K(ζN ). Let M be a prime of OL that divides mOL . We will now show that for each γ ∈ SL2 (Z), the Chinese Remainder Theorem implies that there is an element Aγ ∈ L∗ such that (9.4.3)

Aγ · f [γ]k ∈ Mk (Γ(N ), OL )

and

ordM(Aγ · f [γ]k ) < ∞.

First find A ∈ L∗ such that A · f [γ]k has coefficients in OL . Choose α ∈ M with α 6∈ M2 , and find a negative power αt such that αt · A · f [γ]k has Mintegral coefficients and finite valuation. This is possible because we assumed

173

9.4. Congruences between Newforms

that f is nonzero. Use the Chinese Remainder Theorem to find β ∈ OL such that β ≡ 1 (mod M) and β ≡ 0 (mod ℘) for each prime ℘ 6= M that divides (α). Then for some s we have β s · αt · A · f [γ]k = Aγ · f [γ]k ∈ Mk (Γ(N ), OL ) and ordM(Aγ · f [γ]k ) < ∞. Write

SL2 (Z) =

m [

Γγi

i=1

with γ1 =

( 10 01 ),

and let F =f·

m Y i=2

Aγi · f [γi ]k .

Then F ∈ Mkm (SL2 (Z)) and since M ∩ OK = m, we have ordM(f ) = ordm(f ), so km ordM(F ) ≥ ordM(f ) = ordm(f ) > . 12 Thus we can apply Case 1 to conclude that ordM(F ) = +∞. Thus (9.4.4)

∞ = ordM(F ) = ordm(f ) +

m X

ordM(Aγi f [γ]k ),

i=2

so ordm(f ) = +∞, because of (9.4.3). We next obtain a better bound when f is a cusp form. Since [γ]k preserves cusp forms, ordM(Aγi f [γ]k ) ≥ N1 for each i. Thus m−1 km m−1 = ordm(f ) + > , N N 12 since now we are merely assuming that ordM(F ) ≥ ordM(f ) +

ordm(f ) >

km m − 1 − . 12 N

Thus we again apply Case 1 to conclude that ordM(F ) = +∞, and using (9.4.4), conclude that ordm(f ) = +∞. ¤ Corollary 9.19. Let m be a prime ideal in the ring of integers O of a number field. Suppose f, g ∈ Mk (Γ, O) are modular forms and an (f ) ≡ an (g)

(mod m)

174

9. Computing with Newforms

for all

 km m − 1   − if f − g ∈ Sk (Γ, O), N n ≤ 12   km otherwise, 12 where m = [SL2 (Z) : Γ]. Then f ≡ g (mod m). Buzzard proved the following corollary, which is extremely useful in practical computations. It asserts that the Sturm bound for modular forms with character is the same as the Sturm bound for Γ0 (N ). Corollary 9.20 (Buzzard). Let m be a prime ideal in the ring of integers O of a number field. Suppose f, g ∈ Mk (N, ε, O) are modular forms with Dirichlet character ε : (Z/N Z)∗ → C∗ and assume that an (f ) ≡ an (g) where

(mod m)

for all

n≤

km , 12

¶ Yµ 1 m = [SL2 (Z) : Γ0 (N )] = #P (Z/N Z) = N · 1+ . p 1

p|N

Then f ≡ g (mod m). Proof. Let h = f − g and let r = km/12, so ordm(h) > r. Let s be the order of the Dirichlet character ε. Then hs ∈ Mks (Γ0 (N )) and

ksm . 12 By Theorem 9.18, we have ordm(hs ) = ∞, so ordm(h) = ∞. It follows that f ≡ g (mod m). ¤ ordm(hs ) > sr =

9.4.1.3. Congruence for Newforms. Sturm’s paper [Stu87] also applies some results of Asai on q-expansions at various cusps to obtain a more refined result for newforms. Theorem 9.21 (Sturm). Let N be a positive integer that is square-free, and suppose f and g are two newforms in Sk (N, ε, O), where O is the ring of integers of a number field, and suppose that m is a maximal ideal of O. Let I be an arbitrary subset of the prime divisors of N . If ap (f ) = ap (g) for all p ∈ I and if ap (f ) ≡ ap (g) (mod m) for all primes

p≤ then f ≡ g (mod m).

k · [SL2 (Z) : Γ0 (N )] , 12 · 2#I

175

9.4. Congruences between Newforms

The paper [BS02] contains a similar result about congruences between newforms, which does not require that the level be square-free. Recall from Definition 4.18 that the conductor of a Dirichlet character ε is the largest divisor c of N such that ε factors through (Z/cZ)× . Theorem 9.22. Let N > 4 be any integer, and suppose f and g are two normalized eigenforms in Sk (N, ε; O), where O is the ring of integers of a number field, and suppose that m is a maximal ideal of O. Let I be the set N . If of prime divisors of N that do not divide cond(ε) ap (f ) ≡ ap (g)

(mod m)

for all primes p ∈ I and for all primes p≤ then f ≡ g (mod m).

k · [SL2 (Z) : Γ0 (N )] , 12 · 2#I

For the proof, see Lemma 1.4 and Corollary 1.7 in [BS02, §1.3]. 9.4.2. Generating the Hecke Algebra. The following theorem appeared in [LS02, Appendix], except that we give a better bound here. It is a nice application of the congruence result above, which makes possible explicit computations with Hecke rings T. Theorem 9.23. Suppose Γ is a congruence subgroup that contains Γ1 (N ) and let km m − 1 − , (9.4.5) r= 12 N where m = [SL2 (Z) : Γ]. Then the Hecke algebra T = Z[. . . , Tn , . . .] ⊂ End(Sk (Γ)) is generated as a Z-module by the Hecke operators Tn for n ≤ r. Proof. For any ring R, let Sk (N, R) = Sk (N ; Z) ⊗ R, where Sk (N ; Z) ⊂ Z[[q]] is the submodule of cusp forms with integer Fourier expansion at the cusp ∞, and let TR = T ⊗Z R. For any ring R, there is a perfect pairing Sk (N, R) ⊗R TR → R given by hf, T i 7→ a1 (T (f )) (this is true for R = Z, hence for any R).

Let M be the submodule of T generated by T1 , T2 , . . . , Tr , where r is the largest integer ≤ kN 12 · [SL2 (Z) : Γ]. Consider the exact sequence of additive abelian groups i

0 → M → T → T/M → 0.

176

9. Computing with Newforms

Let p be a prime and use the fact that tensor product is right exact to obtain an exact sequence i

M ⊗ Fp → T ⊗ Fp → (T/M ) ⊗ Fp → 0.

Suppose that f ∈ Sk (N, Fp ) pairs to 0 with each of T1 , . . . , Tr . Then am (f ) = a1 (Tm f ) = hf, Tm i = 0

in Fp for each m ≤ r. By Theorem 9.18, it follows that f = 0. Thus the pairing restricted to the image of M ⊗Fp in TFp is nondegenerate, so because (9.4.5) is perfect, it follows that dimFp i(M ⊗ Fp ) = dimFp Sk (N, Fp ).

Thus (T/M ) ⊗ Fp = 0. Repeating the argument for all primes p shows that T/M = 0, as claimed. ¤ Remark 9.24. In general, the conclusion of Theorem 9.23 is not true if one considers only Tn where n runs over the primes less than the bound. Consider, for example, S2 (11), where the bound is 1 and there are no primes ≤ 1. However, the Hecke algebra is generated as an algebra by operators Tp with p ≤ r.

9.5. Exercises 9.1 Prove that the group Γ1 (N ) is a normal subgroup of Γ0 (N ) and that the quotient Γ0 (N )/Γ1 (N ) is isomorphic to (Z/N Z)∗ . 9.2 Prove that the operators hdi are elements of Z[. . . , Tn , . . .]. [Hint: Use Dirichlet’s theorem on primes in arithmetic progression.] 9.3 Find an example like Example 9.6 but in which the new subspace is nonzero. More precisely, find an integer N such that the Hecke ring on S2 (Γ0 (N )) is not equal to the ring generated by Hecke operators Tn with gcd(n, N ) = 1 and S2 (Γ0 (N ))new 6= 0.

9.4 (a) Following Example 9.15, compute a basis for S2 (Γ0 (31)). (b) Use Algorithm 9.12 to compute a basis for S2 (Γ0 (31)).

Chapter 10

Computing Periods

This chapter is about computing period maps associated to newforms. We assume you have read Chapters 8 and 9 and that you are familiar with abelian varieties at the level of [Ros86]. In Section 10.1 we introduce the period map and give some examples of situations in which computing it is relevant. Section 10.2 is about how to use the period mapping to attach an abelian variety to any newform. In Section 10.3, we introduce extended modular symbols, which are the key computational tool for quickly computing periods of modular symbols. We turn to numerical computation of period integrals in Section 10.4, and in Section 10.5 we explain how to use Atkin-Lehner operators to speed convergence. In Section 10.6 we explain how to compute the full period map with a minimum amount of work. Section 10.7 briefly sketches three approaches to computing all elliptic curves of a given conductor. This chapter was inspired by [Cre97a], which similar algoP contains rithms in the special case of a newform f = an q n ∈ S2 (Γ0 (N )) with an ∈ Z.

See also [Dok04] for algorithmic methods to compute special values of very general L-functions, which can be used for approximating L(f, s) for arbitrary s.

177

178

10. Computing Periods

10.1. The Period Map Let Γ be a subgroup of SL2 (Z) that contains Γ1 (N ) for some N , and suppose X f= an q n ∈ Sk (Γ) n≥1

is a newform (see Definition 9.9). In this chapter we describe how to approximately compute the complex period mapping given by

Φf : Mk (Γ) → C, Φf (P {α, β}) = hf, P {α, β}i =

Z

β

f (z)P (z, 1)dz, α

as in Section 8.5. As an application, we can approximate the special values L(f, j), for j = 1, 2, . . . , k − 1 using (8.5.5). We can also compute the period lattice attached to a modular abelian variety, which is an important step, e.g., in enumeration of Q-curves (see, e.g., [GLQ04]) or computation of a curve whose Jacobian is a modular abelian variety Af (see, e.g., [Wan95]).

10.2. Abelian Varieties Attached to Newforms Fix a newform f ∈ Sk (Γ), where Γ1 (N ) ⊂ Γ for some N . Let f1 , . . . , fd be the Gal(Q/Q)-conjugates of f , where Gal(Q/Q) acts via its action on the Fourier coefficients, which are algebraic integers (since they are the eigenvalues of matrices with integer entries). Let (10.2.1)

Vf = Cf1 ⊕ · · · ⊕ Cfd ⊂ Sk (Γ)

be the subspace of cusp forms spanned by the Gal(Q/Q)-conjugates of f . One can show using the results discussed in Section 9.2 that the above sum is direct, i.e., that Vf has dimension d. The integration pairing induces a T-equivariant homomorphism Φf : Mk (Γ) → Vf∗ = HomC (Vf , C)

from modular symbols to the C-linear dual Vf∗ of Vf . Here T acts on Vf∗ via (ϕt)(x) = ϕ(tx), and this homomorphism is T-stable by Theorem 8.21. The abelian variety attached to f is the quotient Af (C) = Vf∗ /Φf (Sk (Γ; Z)). Here Sk (Γ; Z) = Sk (Γ), and we include the Z in the notation to emphasize that these are integral modular symbols. See [Shi59] for a proof that Af (C) is an abelian variety (in particular, Φf (Sk (Γ; Z)) is a lattice, and Vf∗ is equipped with a nondegenerate Riemann form). When k = 2, we can also construct Af as a quotient of the modular Jacobian Jac(XΓ ), so Af is an abelian variety canonically defined over Q.

10.3. Extended Modular Symbols

179

In general, we have an exact sequence 0 → Ker(Φf ) → Sk (Γ) → Vf∗ → Af (C) → 0. Remark 10.1. When k = 2, the abelian variety Af has a canonical structure of abelian variety over Q. Moreover, there is a conjecture of Ribet and Serre in [Rib92] that describes the simple abelian varieties A over Q that should arise via this construction. In particular, the conjecture is that A is isogenous to some abelian variety Af if and only if End(A/Q) ⊗ Q is a number field of degree dim(A). The abelian varieties Af have this property since Q(. . . , an (f ), . . .) embeds in End(A/Q)⊗Q and the endomorphism ring over Q has degree at most dim(A) (see [Rib92] for details). Ribet proves that his conjecture is a consequence of Serre’s conjecture [Ser87] on modularity of mod p odd irreducible Galois representations (see Section 1.5). Much of Serre’s conjecture has been proved by Khare and Wintenberger (not published). In particular, it is a theorem that if A is a simple abelian variety over Q with End(A/Q) ⊗ Q a number field of degree dim(A) and if A has good reduction at 2, then A is isogenous to some abelian variety Af . Remark 10.2. When k > 2, there is an object called a Grothendieck motive that is attached to f and has a canonical “structure over Q”. See [Sch90].

10.3. Extended Modular Symbols In this section, we extend the notion of modular symbols to allows symbols of the form P {w, z} where w and z are arbitrary elements of h∗ = h ∪ P1 (Q). Definition 10.3 (Extended Modular Symbols). The abelian group Mk of extended modular symbols of weight k is the Z-span of symbols P {w, z}, with P ∈ Vk−2 a homogeneous polynomial of degree k−2 with integer coefficients, modulo the relations P · ({w, y} + {y, z} + {z, w}) = 0 and modulo any torsion. Fix a finite index subgroup Γ ⊂ SL2 (Z). Just as for usual modular symbols, Mk is equipped with an action of Γ, and we define the space of extended modular symbols of weight k for Γ to be the quotient Mk (Γ) = (Mk /hγx − x : γ ∈ Γ, x ∈ Mk i)/ tor . The quotient Mk (Γ) is torsion-free and fixed by Γ. The integration pairing extends naturally to a pairing ¢ ¡ (10.3.1) Sk (Γ) ⊕ S k (Γ) × Mk (Γ) → C,

180

10. Computing Periods

where we recall from (8.5.1) that S k (Γ) denotes the space of antiholomorphic cusp forms. Moreover, if ι : Mk (Γ) → Mk (Γ) is the natural map, then ι respects (10.3.1) in the sense that for all f ∈ Sk (Γ) ⊕ S k (Γ) and x ∈ Mk (Γ), we have hf, xi = hf, ι(x)i. As we will see soon, it is often useful to replace x ∈ Mk (Γ) first by ι(x) P andPthen by an equivalent sum yi of symbols yi ∈ Mk (N, ε) such that hf, yi i is easier to compute numerically than hf, xi. ¡ ¢ Let ε be a Dirichlet character of modulus N . If γ = ac db ∈ SL2 (Z), let ε(γ) = ε(d). Let Mk (N, ε) be the quotient of Mk (N, Z[ε]) by the relations γ(x) − ε(γ)x, for all x ∈ Mk (N, Z[ε]), γ ∈ Γ0 (N ), and modulo any torsion.

10.4. Approximating Period Integrals In this section we assume Γ is a congruence subgroup of SL2 (Z) that contains Γ1 (N ) for some N . Suppose α ∈ h, so Im(α) > 0 and m is an integer such that 0 ≤ m ≤ k − 2, and consider the extended modular symbol X m Y k−2−m {α, ∞}. Let h·, ·i denotePthe integration pairing from Section 8.5. n Given an arbitrary cusp form f = ∞ n=1 an q ∈ Sk (Γ), we have D E (10.4.1) Φf (X m Y k−2−m {α, ∞}) = f, X m Y k−2−m {α, ∞} Z ∞ (10.4.2) f (z)z m dz = α

(10.4.3)

=

∞ X

n=1

an

Z

∞

e2πinz z m dz.

α

The reversal of summation and integration is justified because the imaginary part of α is positive so that the sum converges absolutely. The following lemma is useful for computing the above infinite sum. Lemma 10.4.  Z ∞ m s m−s X  (−1) α e2πinz z m dz = e2πinα (10.4.4) (2πin)s+1 α s=0

Proof. See Exercise 10.1

m Y

j=(m+1)−s



j .

¤

In practice we will usually be interested in computing the period map Φf when f ∈ Sk (Γ) is a newform. Since f is a newform, there is a Dirichlet character ε such that f ∈ Sk (N, ε). The period map Φf : Mk (Γ) → C then

10.4. Approximating Period Integrals

181

factors through the quotient Mk (N, ε), so it suffices to compute the period map on modular symbols in Mk (N, ε). The following proposition is an analogue of [Cre97a, Prop. 2.1.1(5)]. Proposition 10.5. For any γ ∈ Γ0 (N ), P ∈ Vk−2 and α ∈ h∗ , we have the following relation in Mk (N, ε): (10.4.5) P {∞, γ(∞)} = P {α, γ(α)} + (P − ε(γ)γ −1 P ){∞, α}

(10.4.6)

= ε(γ)(γ −1 P ){α, ∞} − P {γ(α), ∞}.

Proof. By definition, if x ∈ Mk (N, ε) is a modular symbol and γ ∈ Γ0 (N ), then γx = ε(γ)x. Thus ε(γ)γ −1 x = x, so P {∞, γ(∞)} = P {∞, α} + P {α, γ(α)} + P {γ(α), γ(∞)}

= P {∞, α} + P {α, γ(α)} + ε(γ)γ −1 (P {γ(α), γ(∞)})

= P {∞, α} + P {α, γ(α)} + ε(γ)(γ −1 P ){α, ∞}

= P {α, γ(α)} + P {∞, α} − ε(γ)(γ −1 P ){∞, α}

= P {α, γ(α)} + (P − ε(γ)γ −1 P ){∞, α}.

The second equality in the statement of the proposition now follows easily. ¤ In the case of weight 2 and trivial character, the “error term” (10.4.7)

(P − ε(γ)γ −1 P ){∞, α}

vanishes since P is constant and ε(γ) = 1. In general this term does not vanish. However, we can suitably modify the formulas found in [Cre97a, 2.10] and still obtain an algorithm for computing period integrals. Algorithm 10.6 (Period Integrals). Given γ ∈ Γ0 (N ), P ∈ Vk−2 and f ∈ Sk (N, ε) presented as a q-expansion to some precision, this algorithm outputs an approximation to the period integral hf, P {∞, γ(∞)}i. ¢ ¡ (1) Write γ = cNa db ∈ Γ0 (N ), with a, b, c, d ∈ Z, and set α = −d+i cN in Proposition 10.5. (2) Replacing γ by −γ if necessary, we find that the imaginary parts 1 of α and γ(α) = a+i cN are both equal to the positive number cN . (3) Use (10.4.3) and Lemma 10.4 to compute the integrals that appear in Proposition 10.5. It would be nice if the modular symbols of the form P {∞, γ(∞)} for P ∈ Vk−2 and γ ∈ Γ0 (N ) were to generate a large subspace of Mk (N, ε) ⊗ Q. When k = 2 and ε = 1, Manin proved in [Man72] that the map Γ0 (N ) → H1 (X0 (N ), Z) sending γ to {0, γ(0)} is a surjective group homomorphism. When k > 2, the author does not know a similar group-theoretic statement. However, we have the following theorem.

182

10. Computing Periods -Q βα r

P α r

r

γα Q -

P r

r

r

∞

γ∞

β∞

Figure 10.4.1. “Transporting” a transportable modular symbol.

Theorem 10.7. Any element of Sk (N, ε) can be written in the form

n X i=1

Pi {∞, γi (∞)}

for some P Pi ∈ PVk−2 and γi ∈ Γ0 (N ). Moreover, Pi and γi can be chosen so that Pi = ε(γi )γi−1 (Pi ), so the error term (10.4.7) vanishes. The author and Helena Verrill prove this theorem in [SV01]. The condition that the error term vanishes means that one can replace ∞ by any α in the expression for the modular symbol and obtain an equivalent modular symbol. For this reason, we call such modular symbols transportable, as illustrated in Figure 10.4.1. Note that in general not every element of the form P {∞, γ(∞)} must lie in Sk (N, ε). However, if γP = P , then P {∞, γ(∞)} does lie in Sk (N, ε). It would be interesting to know under what circumstances Sk (N, ε) is generated by symbols of the form P {∞, γ(∞)} with γP = P . This sometimes fails for k odd; for example, when k = 3, the condition γP = P implies that γ ∈ Γ0 (N ) has an eigenvector with eigenvalue 1, and hence is of finite order. When k is even, the author can see no obstruction to generating Sk (N, ε) using such symbols.

10.5. Speeding Convergence Using Atkin-Lehner

183

10.5. Speeding Convergence Using Atkin-Lehner ¡ ¢ Let wN = N0 −10 ∈ Mat2 (Z). Consider the Atkin-Lehner involution WN on Mk (Γ1 (N )), which is defined by WN (f ) = N (2−k)/2 · f |[wN ]k µ ¶ 1 (2−k)/2 =N ·f − · N k−1 · (N z)−k Nz ¶ µ 1 −k/2 −k . =N ·z ·f − Nz Here we take the positive square root if k is odd. Then WN2 = (−1)k is an involution when k is even. There is an operator on modular symbols, which we also denote WN , which is given by WN (P {α, β}) = N (2−k)/2 · wN (P ){wN (α), wN (β)} ½ ¾ 1 1 (2−k)/2 =N · P (−Y, N X) − ,− , αN βN and one has that if f ∈ Sk (Γ1 (N )) and x ∈ Mk (Γ1 (N )), then hWN (f ), xi = hf, WN (x)i. If ε is a Dirichlet character of modulus N , then the operator WN sends Sk (N, ε) to Sk (Γ1 (N ), ε). Thus if ε2 = 1, then WN preserves Sk (N, ε). In particular, WN acts on Sk (Γ0 (N )). The next proposition shows how to compute the pairing hf, P {∞, γ(∞)}i under certain restrictive assumptions. It generalizes a result of [Cre97b] to higher weight. Proposition 10.8. Let f ∈ Sk (N, ε) be a cusp form which is an eigenform for the Atkin-Lehner operator WN having eigenvalue w ∈ {±1} (thus ε2 = 1 and k is even). Then for any γ ∈ Γ0 (N ) and any P ∈ Vk−2 , with the property that γP = ε(γ)P , we have the following formula, valid for any α ∈ h: D P (Y, −N X) hf, P {∞, γ(∞)}i = f, w {wN (α), ∞} N k/2−1 ¶ µ o E P (Y, −N X) n √ N , ∞ − P {γ(α), ∞} . i/ + P −w N k/2−1 Here wN (α) = −

1 . Nα

Proof. By Proposition 10.5 our condition on P implies that P {∞, γ(∞)} = P {α, γ(α)}. We describe the steps of the following computation below.

184

10. Computing Periods

D

E f, P {α, γ(α)} D E √ √ = f, P {α, i/ N } + P {i/ N , W (α)} + P {W (α), γ(α)} ¿ À √ √ W (P ) = f, w k/2−1 {W (α), i/ N } + P {i/ N , W (α)} + P {W (α), γ(α)} . N

For the first equality, we break the path into three paths, and in the second, we apply the W -involution to the first term and use that the action of W is compatible with the pairing h , i and that f is an eigenvector with eigenvalue w. In the following sequence of equalities we combine √ the first two terms N } by {W (α), ∞} + and break up the third; then we replace {W (α), i/ √ {∞, i/ N } and regroup: √ √ W (P ) N } + P {i/ N , W (α)} + P {W (α), γ(α)} {W (α), i/ Nµk/2−1 ¶ √ W (P ) = w k/2−1 − P {W (α), i/ N } + P {W (α), ∞} − P {γ(α), ∞} N ¶ µ √ W (P ) W (P ) = w k/2−1 {W (α), ∞} + P − w k/2−1 {i/ N , ∞} − P {γ(α), ∞}. N N

w

¤ ³

´

A good choice for α is α = γ −1 db + d√i N , so that W (α) = dc + d√i N . This maximizes the minimum of the imaginary parts of α and W (α), which results in series that converge more quickly. ¡ ¢ Let γ = ac db ∈ Γ0 (N ). The polynomial P (X, Y ) = (cX 2 + (d − a)XY − bY 2 )

k−2 2

satisfies γ(P ) = P . We obtained this formula by viewing Vk−2 as the (k−2)th symmetric product of the 2-dimensional space on which Γ0 (N ) acts naturally. For example, observe that since det(γ) = 1, the symmetric product of two eigenvectors for γ is an eigenvector in V2 having eigenvalue 1. For the same reason, if ε(γ) 6= 1, there need not be a polynomial P (X, Y ) such that γ(P ) = ε(γ)P . One remedy is to choose another γ so that ε(γ) = 1. √ Since the imaginary parts of the terms i/ N , α and W (α) in the proposition are all relatively large, the sums appearing at the beginning of Section 10.4 converge quickly if d is small. It is important to choose γ in Proposition 10.8 with d small; otherwise the series will converge very slowly. Remark 10.9. Is there a generalization of Proposition 10.8 without the restrictions that ε2 = 1 and k is even?

185

10.6. Computing the Period Mapping

10.5.1. Another Atkin-Lehner Trick. Suppose E is an elliptic curve and let L(E, s) be the corresponding L-function. Let ε ∈ {±1} be the root number of E, i.e., the sign of the functional equation for L(E, s), so Λ(E, s) = εΛ(E, 2−s), where Λ(E, s) = N s/2 (2π)−s Γ(s)L(E, s). Let f = fE be the modular form associated to E (which exists by [Wil95, BCDT01]). If WN (f ) = wf , then ε = −w (see Exercise 10.2). We have L(E, 1) = −2π

Z

∞

f (z) dz

0

= −2πi hf, {0, ∞}i D E √ √ = −2πi f, {0, i/ N } + {i/ N , ∞} E D √ √ = −2πi wf, {wN (0), wN (i/ N )} + {i/ N , ∞} D E √ √ = −2πi wf, {∞, i/ N } + {i/ N , ∞} D √ E = −2πi (w − 1) f, {∞, i/ N } . If w = 1, then L(E, 1) = 0. If w = −1, then (10.5.1)

∞ X √ E an −2πn/√N . e L(E, 1) = 4πi f, {∞, i/ N } = 2 n

D

n=1

For more about computing with L-functions of elliptic curves, including a trick for computing ε quickly without directly computing WN , see [Coh93, §7.5] and [Cre97a, §2.11]. One can also find higher derivatives L(r) (E, 1) by a formula similar to (10.5.1) (see [Cre97a, §2.13]). The methods in this chapter for obtaining rapidly converging series are not just of computational interest; see, e.g., [Gre83] for a nontrivial theoretical application to the Birch and Swinnerton-Dyer conjecture.

10.6. Computing the Period Mapping Fix a newform f = be as in (10.2.1).

P

an q n ∈ Sk (Γ), where Γ1 (N ) ⊂ Γ for some N . Let Vf

Let Θf : Mk (Γ; Q) → V be any Q-linear map with the same kernel as Φf ; we call any such map a rational period mapping associated to f . Let Φf be the period mapping associated to the Gal(Q/Q)-conjugates of f . We

186

10. Computing Periods

have a commutative diagram Mk (Γ; Q)

II II II Θf III $$

Φf

V

// HomC (Vf , C) 99 if rrr r r rrr + ® rrr

Recall from Section 10.2 that the cokernel of Φf is the abelian variety Af (C). The Hecke algebra T acts on the linear dual Mk (Γ; Q)∗ = Hom(Mk (Γ), Q) by (tϕ)(x) = ϕ(tx). Let I = If ⊂ T be the kernel of the ring homomorphism T → Z[a2 , a3 , . . .] that sends Tn to an . Let Mk (Γ; Q)∗ [I] = {ϕ ∈ Mk (Γ; Q)∗ : tϕ = 0 all t ∈ I}.

Since f is a newform, one can show that Mk (Γ; Q)∗ [I] has dimension d. Let θ1 , . . . , θd be a basis for Mk (Γ; Q)∗ [I], so Ker(Φf ) = Ker(θ1 ) ⊕ · · · ⊕ Ker(θd ). We can thus compute Ker(Φf ), hence a choice of Θf . To compute Φf , it remains to compute if . Let Sk (Γ; Q) denote the space of cusp forms with q-expansion in Q[[q]]. By Exercise 10.3 Sk (Γ; Q)[I] = Sk (Γ)[I] ∩ Q[[q]]

is a Q-vector space of dimension d. Let g1 , . . . , gd be a basis for this Q-vector space. We will compute Φf with respect to the basis of HomQ (Sk (Γ; Q)[I]; C) dual to this basis. Choose elements x1 , . . . , xd ∈ Mk (Γ) with the following properties: (1) Using Proposition 10.5 or Proposition 10.8, it is possible to compute the period integrals hgi , xj i, i, j ∈ {1, . . . , d}, efficiently.

(2) The 2d elements v + η(v) and v − η(v) for v = Θf (x1 ), . . . , Θf (xd ) span a space of dimension 2d (i.e., they span Mk (Γ)/ Ker(Φf )). Given this data, we can compute if (v + η(v)) = 2Re(hg1 , xi i, . . . , hgd , xi i) and if (v − η(v)) = 2iIm(hg1 , xi i, . . . , hgd , xi i).

We break the integrals into real and imaginary parts because this increases the precision of our answers. Since the vectors vn + η(vn ) and vn − η(vn ), n = 1, . . . , d, span Mk (N, ε; Q)/ Ker(Φf ), we have computed if .

10.7. All Elliptic Curves of Given Conductor

187

Remark 10.10. We want to find symbols xi satisfying the conditions of Proposition 10.8. This is usually possible when d is very small, but in practice it is difficult when d is large. Remark 10.11. The above strategy was motivated by [Cre97a, §2.10].

10.7. All Elliptic Curves of Given Conductor Using modular symbols and the period map, we can compute all elliptic curves over Q of conductor N , up to isogeny. The algorithm in this section gives all modular elliptic curves (up to isogeny), i.e., elliptic curves attached to modular forms, of conductor N . Fortunately, it is now known by [Wil95, BCDT01, TW95] that every elliptic curve over Q is modular, so the procedure of this section gives all elliptic curves (up to isogeny) of given conductor. See [Cre06] for a nice historical discussion of this problem. Algorithm 10.12 (Elliptic Curves of Conductor N ). Given N > 0, this algorithm outputs equations for all elliptic curves of conductor N , up to isogeny. (1) [Modular Symbols] Compute M2 (Γ0 (N )) using Section 8.7. (2) [Find Rational Eigenspaces] Find the 2-dimensional eigenspaces V in M2 (Γ0 (N ))new that correspond to elliptic curves. Do not use the algorithm for decomposition from Section 7.5, which is too complicated and gives more information than we need. Instead, for the first few primes p ∤ N , compute all eigenspaces Ker(Tp − a), √ √ where a runs through integers with −2 p < a < 2 p. Intersect these eigenspaces to find the eigenspaces that correspond to elliptic curves. To find just the new ones, either compute the degeneracy maps to lower level or find all the rational eigenspaces of all levels that strictly divide N and exclude them. (3) [Find Newforms] Use 9.14 to compute to some precision P∞Algorithm n each newform f = n=1 an q ∈ Z[[q]] associated to each eigenspace V found in step (2). (4) [Find Each Curve] For each newform f found in step (3), do the following: (a) [Period Lattice] Compute the corresponding period lattice Λ = Zω1 + Zω2 by computing the image of Φf , as described in Section 10.6. (b) [Compute τ ] Let τ = ω1 /ω2 . If Im(τ ) < 0, swap ω1 and ω2 , so Im(τ ) > 0. By successively applying generators of SL2 (Z), we find an SL2 (Z) equivalent element τ ′ in F, i.e., |Re(τ ′ )| ≤ 1/2 and |τ | ≥ 1.

188

10. Computing Periods

(c) [c-invariants] Compute the invariants c4 and c6 of the lattice Λ using the following rapidly convergent series: ! µ ¶4 Ã ∞ X 2π n3 q n c4 = , · 1 + 240 ω2 1 − qn n=1 ! µ ¶6 Ã ∞ X n5 q n 2π , · 1 − 504 c6 = ω2 1 − qn n=1

′ e2πiτ ,

τ′

where q = where is as in step (4b). A theorem of Edixhoven (that the Manin constant is an integer) implies that the invariants c4 and c6 of Λ are integers, so it is only necessary to compute Λ to large precision to completely determine them. (d) [Elliptic Curve] An elliptic curve with invariants c4 and c6 is c4 c6 E : y 2 = x3 − x − . 48 864 (e) [Prove Correctness] Using Tate’s algorithm, find the conductor of E. If the conductor is not N , then recompute c4 and c6 using more terms of f and real numbers to larger precision, etc. If the conductor is N , compute the coefficients bp of the modular form g = gE attached to the elliptic curve E, for p ≤ #P1 (Z/N Z)/6. Verify that ap = bp , where ap are the coefficients of f . If this equality holds, then E must be isogenous to the elliptic curve attached to f , by the Sturm bound (Theorem 9.18) and Faltings’s isogeny theorem. If the equality fails for some p, recompute c4 and c6 to larger precision. There are numerous tricks to optimize the above algorithm. For example, − often one can work separately with Mk (Γ0 (N ))+ new and Mk (Γ0 (N ))new and get enough information to find E, up to isogeny (see [Cre97b]). Once we have one curve from each isogeny class of curves of conductor N , we find each curve in each isogeny class (which is another interesting problem discussed in [Cre97a]), hence all curves of conductor N . If E/Q is an elliptic curve, then any curve isogenous to E is isogenous via a chain of isogenies of prime degree. There is an a priori bound on the degrees of these isogenies due to Mazur. Also, there are various methods for finding all isogenies of a given degree with domain E. See [Cre97a, §3.8] for more details. 10.7.1. Finding Curves: S-Integral Points. In this section we briefly survey an alternative approach to finding curves of a given conductor by finding integral points on other elliptic curves. Cremona and others have developed a complementary approach to the problem of computing all elliptic curves of given conductor (see [CL04]).

10.7. All Elliptic Curves of Given Conductor

189

Instead of computing all curves of given conductor, we instead consider the seemingly more difficult problem of finding all curves with good reduction outside a finite set S of primes. Since one can compute the conductor of a curve using Tate’s algorithm [Tat75, Cre97a, §3.2], if we know all curves with good reduction outside S, we can find all curves of conductor N by letting S be the set of prime divisors of N . There is a strategy for finding all curves with good reduction outside S. It is not an algorithm, in the sense that it is always guaranteed to terminate (the modular symbols method above is an algorithm), but in practice it often works. Also, this strategy makes sense over any number field, whereas the modular symbols method does not (there are generalizations of modular symbols to other number fields). Fix a finite set S of primes of a number field K. It is a theorem of Shafarevich that there are only finitely many elliptic curves with good reduction outside S (see [Sil92, Section IX.6]). His proof uses that the group of S-units in K is finite and Siegel’s theorem that there are only finitely many S-integral points on an elliptic curve. One can make all this explicit, and sometimes in practice one can compute all these S-integral points. The problem of finding all elliptic curves with good reduction outside of S can be broken into several subproblems, the main ones being (1) determine the following finite subgroup of K ∗ /(K ∗ )m : K(S, m) = {x ∈ K ∗ /(K ∗ )m : m | ordp(x) all p 6∈ S}; (2) find all S-integral points on certain elliptic curves y 2 = x3 + k. In [CL04], there is one example, where they find all curves of conductor N = 28 · 172 = 73984 by finding all curves with good reduction outside {2, 17}. They finds 32 curves of conductor 73984 that divide into 16 isogeny classes. (Note that dim S2 (Γ0 (N )) = 9577.) 10.7.2. Finding Curves: Enumeration. One can also find curves by simply enumerating Weierstrass equations. For example, the paper [SW02] discusses a database that the author and Watkins created that contains hundreds of millions of elliptic curves. It was constructed by enumerating Weierstrass equations of a certain form. This database does not contain every curve of each conductor included in the database. It is, however, fairly complete in some cases. For example, using the Mestre method of graphs [Mes86], we verified in [JBS03] that the database contains all elliptic curve of prime conductor < 234446, which implies that the smallest conductor rank 4 curve is composite.

190

10. Computing Periods

10.8. Exercises 10.1 Prove Lemma 10.4. 10.2 Suppose f ∈ S2 (Γ0 (N )) is a newform and that WN (f ) = wf . Let Λ(E, s) = N s/2 (2π)−s Γ(s)L(E, s). Prove that Λ(E, s) = −wΛ(E, 2 − s). √ R [Hint: Show that Λ(f, s) = 0,∞ f (iy/ N )y s−1 dy. Then substitute 1/y for y.] P 10.3 Let f = an q n ∈ C[[q]] be a power series whose coefficients an together generate a number field K of degree d over Q. Let Vf be the complex vector space spanned by the Gal(Q/Q)-conjugates of f . (a) Give an example to show that Vf need not have dimension d. (b) Suppose Vf has dimension d. Prove that Vf ∩ Q[[q]] is a Qvector space of dimension d. 10.4 Find an elliptic curve of conductor 11 using Section 10.7.

Chapter 11

Solutions to Selected Exercises

11.1. Chapter 1 ¢ ¡ ∈ GL2 (R) is a matrix with (1) Exercise 1.1. Suppose γ = ac nb d positive determinant. Then γ is a linear fractional transformation that fixes the real line, so it must either fix or swap the upper and lower half planes. Now

γ(i) =

ac + bd + (ad − bc)i ai + b = , ci + d d 2 + c2

so since det γ = ad − bc > 0, the imaginary part of γ(i) is positive; hence γ sends the upper half plane to itself. (2) Exercise 1.2. Avoiding poles, the quotient rule for differentiation goes through exactly as in the real case, so any rational function f (z) = p(z)/q(z) (p, q ∈ C[z]) is holomorphic on C−{α : q(α) = 0}. By the fundamental theorem of algebra, this set of poles is finite, and hence it is discrete. Write q(z) = an (z − α1 )r1 · · · (z − αk )rk for each αi and let qi (z) = q(z)/(z − αi )ri which is a polynomial nonzero at αi . Thus for each i we have (z − αi )ri f (z) = p(z)/q ′ (z) is holomorphic at αi and hence f (z) is meromorphic on C. (3) Exercise 1.3. (a) The product f g of two meromorphic functions on the upper half plane is itself meromorphic. Also, for all γ ∈ SL2 (Z) we 191

192

11. Solutions to Selected Exercises

have 1 ((f g) ◦ γ) (cz + d)k+j 1 1 (f ◦ γ) (g ◦ γ) = f g, = k (cz + d) (cz + d)j

(f g)[γ]k+j =

so f g is weakly modular. (b) If f is meromorphic on the upper half plane, then so is 1/f . Now 1 1 1 [γ]−k k = = (cz + d) ((1/f ) ◦ γ) = , f (cz + d)−k f ◦ γ f

so 1/f is a weakly modular form of weight −k. (c) Let f and g be modular functions. Then, as P above, f g is a P ∞ n n and a q weakly modular function. Let ∞ n=m′ bn q be n=m n their q-expansions around any α ∈ P1 (Q); then their formal product is the q-expansion of f g. But the formal product of two Laurent series about the same point is itself a Laurent series with convergence in the intersection of the convergent domains of the original series, so f g has a meromorphic qexpansion at each α ∈ P1 (Q) and hence at each cusp. (d) We are in exactly the same case as in part (c), but because f and g are modular functions, m, m′ ≥ 0 and hence the function is holomorphic at each of its cusps. (4) Exercise 1.4. ¡Let f ¢be a weakly modular function of odd weight 0 −k k. Since γ = −1 0 −1 ∈ SL2 (Z), we have f (z) = (−1) f (γ(z)) = −f (z) so f = 0.

(5) Exercise 1.5. Because SL2 (Z/1Z) is the trivial group, Γ(1) = ker(SL2 (Z) → SL2 (Z/1Z)) must be all of SL2 (Z). As SL2 (Z) = Γ(1) ⊂ Γ1 (1) ⊂ Γ0 (1) ⊂ SL2 (Z), we must have Γ(1) = Γ1 (1) = Γ0 (1) = SL2 (Z).

(6) Exercise 1.6. (a) The group Γ1 (N ) is the inverse image of the subgroup of SL2 (Z/N Z) generated by ( 10 11 ), and the inverse image of a group (under a group homomorphism) is a group. (b) The group contains the kernel of the homomorphism SL2 (Z) → SL2 (Z/N Z), and that kernel has finite index since the quotient is contained in SL2 (Z/N Z), which is finite. (c) Same argument as previous part. (d) The level is at most N since ¡both¢groups contain Γ(N ). It can be no greater than N since 10 N1 is in both groups. (7) Exercise 1.7. See [DS05, Lemma 1.2.2].

193

11.2. Chapter 2

(8) Exercise 1.8. Let α = p/q ∈ Q, where p and q are relatively prime. By the Euclidean ¡algorithm, we can find x, y ∈ Z such that px + ¢ p −y qy = 1. Let γα = q x . Note that γα ∈ SL2 (Z) and γα (∞) = α. Also let γ∞ be the identity map on P1 (Q). Now γβ−1 sends β to ∞ so we have γα ◦ γβ−1 which sends α to β.

11.2. Chapter 2 (1) Exercise 2.1. We have ζ(26) =

1315862 · π 26 . 11094481976030578125

Variation: Compute ζ(28). (2) Exercise 2.2. Omitted. (3) Exercise 2.3. E8 = − =

∞

X B8 σ7 (n)q n +q+ 16 n=2

1 + q + 129q 2 + 2188q 3 + · · · . 480

Variation: Compute E10 . (4) Exercise 2.4. Omitted. (5) Exercise 2.5. We have d = dim S28 = 2. A choice of a, b with 4a + 6b ≤ 14 and 4a + 6b ≡ 4 (mod 12) is a = 1, b = 0. A basis for S28 is then 2(2−1)+0

g1 = ∆F6

F4 = q − 792q 2 − 324q 3 + 67590208q 4 + · · · ,

2(2−2)+0

g2 = ∆2 F6

F4 = q 2 + 192q 3 − 8280q 4 + · · · .

The Victor Miller basis is then f1 = g1 + 729g2 = q + 151740q 3 + 61032448q 4 + · · · ,

f2 = g2 = q 2 + 192q 3 − 8280q 4 + · · · .

Variation: Compute the Victor Miller basis for S30 .

194

11. Solutions to Selected Exercises

(6) Exercise 2.6. From the previous exercise we have f = ∆2 F4 . Then ¶2 µ 3 F4 − F62 · F4 f = ∆2 F4 = −1728 ³ ´3 ³ ´2 2 8 12 µ ¶ − − − E E B4 4 B6 6 8   =  · − E4 −1728 B4 = 5186160E4 E64 − 564480000E44 E62 + 15360000000E47 .

(7) Exercise 2.7. No, it is not always integral. For example, for k = 12, the coefficient of q is −2 · 12/B12 = 65520/691 6∈ Z. Variation: Find, with proof, the set of all k such that the normalized series Fk is integral (use that Bk is eventually very large compared to 2k). (8) Exercise 2.8. We compute the Victor Miller basis to precision great enough to determine T2 . This means we need up to O(q 5 ). f0 = 1 + 2611200q 3 + 19524758400q 4 + · · · , f1 = q + 50220q 3 + 87866368q 4 + · · · ,

f2 = q 2 + 432q 3 + 39960q 4 + · · · . Then the matrix of T2 on  2147483649  0 0

this basis is  0 19524758400 0 2235350016  . 1 39960

(The rows of this matrix are the linear combinations that give the images of the fi under T2 .) This matrix has characteristic polynomial (x − 2147483649) · (x2 − 39960x − 2235350016).

11.3. Chapter 3

¡ ¢ (1) Exercise 3.1. Write g = ac db , so λ′ = aλ+b cλ+d . Let f be the isomor′ phism C/Λ → C/Λ given by f (z) = z/(cλ + d). We have µ ¶ 1 a c aλ + b ∼ a 1 f = − · (mod Z + Zλ′ ), = = N N (cλ + d) N N cλ + d N where the second equality can be verified easily by expanding out each side, and for the congruence we use that N | c. Thus the subgroup of C/Λ generated by N1 is taken isomorphically to the subgroup of C/Λ′ generated by N1 .

195

11.3. Chapter 3

(2) Exercise 3.2. For any integer r, we have ( 10 1r ) ∈ Γ0 (N ), so {0, ∞} = {r, ∞}. Thus 0 = {0, ∞} − {0, ∞} = {n, ∞} − {m, ∞} = {n, ∞} + {∞, m} = {n, m}. (3) Exercise 3.3. (a) (0 : 1), (1 : 0), (1 : 1), . . . , (1, p − 1). (b) p + 1. (c) See [Cre97a, Prop. 2.2.1]. (4) Exercise 3.4. We start with b = 4, a = 7. Then 4 · 2 ≡ 1 (mod 7). Let δ1 = ( 47 12 ) ∈ SL2 (Z). Since δ1 ∈ Γ0 (7), we use the right coset representative ( 10 01 ) and see that {0, 4/7} = {0, 1/2} + ( 10 01 ) {0, ∞}. Repeating process, we have δ2 = ( 12 10 ), which is in the same ¢ ¡ 0 the coset at 1 −10 . Thus {0, 1/2} = ( 01 60 ) {0, ∞} + {0, 0}.

Putting it together gives {0, 4/7} = ( 10 01 ) {0, ∞} + ( 01 60 ) {0, ∞} = [(0, 1)] + [(1, 0)]. (5) Exercise 3.5. (a) Coset representatives for Γ0 (3) in SL2 (Z) are ¶ ¶ µ ¶ µ ¶ µ µ 0 −1 1 0 1 0 1 0 , , , , 1 0 2 1 1 1 0 1 which we refer to below as [r0 ], [r1 ], [r2 ], and [r3 ], respectively. (b) In terms of representatives we have [r0 ] + [r3 ] = 0, [r1 ] + [r2 ] = 0, [r2 ] + [r1 ] = 0, [r3 ] + [r0 ] = 0,

[r0 ] + [r3 ] + [r2 ] = 0, [r1 ] + [r1 ] + [r1 ] = 0, [r2 ] + [r0 ] + [r3 ] = 0, [r3 ] + [r2 ] + [r0 ] = 0.

(c) By the first three relations we have [r2 ] = [r1 ] = 0 = 0[r0 ] and [r3 ] = −1[r0 ]. (d) T2 ([r0 ]) = [r0 ] ( 10 02 ) + [r0 ] ( 20 01 ) + [r0 ] ( 20 11 ) + [r0 ] ( 11 02 ) = [( 10 02 )] + [( 20 01 )] + [( 20 11 )] + [( 11 02 )] = [r0 ] + [r0 ] + [r0 ] + [r2 ] = 3[r0 ].

196

11. Solutions to Selected Exercises

11.4. Chapter 4 (1) Exercise 4.1. Suppose f is a Dirichlet character with modulus N . Then −1 = f (−1) = f (−1 + N ) = 1, a contradiction.

(2) Exercise 4.2. (a) Any finite subgroup of the multiplicative group of a field is cyclic (since the number of roots of a polynomial over a field is at most its degree), so (Z/pZ)∗ is cyclic. Let g be an integer that reduces to a generator of (Z/pZ)∗ . Let x = 1 + p ∈ (Z/pn Z)∗ ; by the binomial theorem xp

n−2

= 1 + pn−2 · p + · · · ≡ 1 + pn−1 6≡ 0 (mod pn ),

so x has order pn−1 . Since p is odd, gcd(pn−1 , p − 1) = 1, so xg has order pn−1 · (p − 1) = ϕ(pn ); hence (Z/pn Z)∗ is cyclic. n−3 (b) By the binomial theorem (1 + 22 )2 6≡ 1 (mod 2n ), so 5 has n−2 n ∗ order 2 in (Z/2 Z) , and clearly −1 has order 2. Since 5 ≡ 1 (mod 4), −1 is not a power of 5 in (Z/2n Z)∗ . Thus the subgroups h−1i and h5i have trivial intersection. The product of their orders is 2n−1 = ϕ(2n ) = #(Z/2n Z)∗ , so the claim follows. Q ei (3) Exercise 4.3. Write n = pi . The order of g divides n, so the condition implies that pei i divides the order of g for each i. Thus the order of g is divisible by the least common multiple of the pei i , i.e., by n. (4) Exercise 4.4. (a) The bijection given by 1 + pn−1 a (mod pn ) 7→ a (mod p) is a homomorphism since (1 + pn−1 a)(1 + pn−1 b) ≡ 1 + pn−1 (a + b)

(mod pn ).

(b) We have an exact sequence 1 → 1 + pZ/pn Z → (Z/pn Z)∗ → (Z/pZ)∗ → 1, so it suffices to solve the discrete log problem in the kernel and cokernel. We prove by induction on n that we can solve the discrete log problem in the kernel easily (compared to known methods for solving the discrete log problem in (Z/pZ)∗ ). We have an exact sequence 1 → 1 + pn−1 Z/pn Z → (Z/pn Z)∗ → (Z/pn−1 Z)∗ → 1. The first part of this problem shows that we can solve the discrete log problem in the kernel, and by induction we can solve it in the cokernel. This completes the proof.

197

11.6. Chapter 6

(5) Exercise 4.5. If ε(5) = 1, then since ε is nontrivial, Exercise 4.2 implies that ε factors through (Z/4Z)∗ , hence has conductor 4 = 21+1 , as claimed. If ε(5) 6= 1, then again from Exercise 4.2 we see that if ε has order r, then ε factors through (Z/2r+2 Z)∗ but nothing smaller. (6) Exercise 4.6. (a) Take f = x2 + 2. (b) The element 2 has order 4. (c) A minimal generator for (Z/25Z)∗ is 2, and the characters are [1], [2], [3], [4]. (d) Each of the four Galois orbits has size 1.

11.5. Chapter 5 (1) Exercise 5.1. The eigenspace Eλ of A with eigenvalue λ is preserved by B, since if v ∈ Eλ , then ABv = BAv = B(λv) = λBv. Because B is diagonalizable, its minimal polynomial equals its characteristic polynomial; hence the same is true for the restriction of B to Eλ , i.e., the restriction of B is diagonalizable. Choose basis for all Eλ so that the restrictions of B to these eigenspaces is diagonal with respect to these bases. Then the concatenation of these bases is a basis that simultaneously diagonalizes A and B. (2) Exercise 5.2. When ε is the trivial character, the Bk,ε are defined by 1 X ε(a)xeax a=1

ex − 1

=

∞

X xk xex x B = x + = . k,ε ex − 1 ex − 1 k! k=0

Thus B1,ε = 1 + B1 = 12 , and for k > 1, we have Bk,ε = Bk . (3) Exercise 5.3. Omitted. (4) Exercise 5.4. The Eisenstein series in our basis for E3 (Γ1 (13)) are of the form E3,1,ε or E3,ε,1 with ε(−1) = (−1)3 = −1. There are six characters ε with modulus 13 such that ε(−1) = −1, and we have the two series E3,1,ε and E3,ε,1 associated to each of these. This gives a dimension of 12.

11.6. Chapter 6 (1) Exercise 6.1.

198

11. Solutions to Selected Exercises

(a) By Proposition 3.10, we have [SL2 (Z) : Γ0 (N )] = #P1 (Z/N Z). By the Chinese Remainder Theorem, Y #P1 (Z/N Z) = #P1 (Z/pordp (N ) Z). p|N

So we are reduced to computing #P1 (Z/pordp (N ) Z). We have (a, b) ∈ (Z/pn Z)2 with gcd(a, b, p) = 1 if and only if (a, b) 6∈ (pZ/pn Z)2 , so there are p2n − p2(n−1) such pairs. The unit group (Z/pn Z)∗ has order ϕ(pn ) = pn − pn−1 . It follows that #P1 (Z/N Z) = (b) Omitted.

p2n − p2(n−1) = pn + pn−1 . pn − pn−1

(2) Exercise 6.2. Omitted. (3) Exercise 6.3. Omitted. (4) Exercise 6.4. Omitted. (5) Exercise 6.5. See the source code to SAGE.

11.7. Chapter 7 (1) Exercise 7.1. Take a basis of W and let G be the matrix whose rows are these basis elements. Let B be the row echelon form of G. After a permutation p of columns, we may write B = pi (I|C), where I is the identity matrix. The matrix A = p−1 (−C t |I), where I is a different sized identity matrix, has the property that W = Ker(A). (2) Exercise 7.2. The answer is no. For example if A = nI is n times the identity matrix and if p | n, then rref(A (mod p)) = 0 but rref(A) (mod p) = I. Q (3) Exercise 7.3. Let T = Ei be an invertible matrix such that T A = E is in (reduced) echelon form and the Ei are elementary matrices, i.e., the result of applying an elementary row operation to the identity matrix. If p is a prime that does not divide any of the nonzero numerators or denominators of the entries of A and any Ei , then rref(A (mod p)) = rref(A) (mod p). This is because E (mod p) is in echelon form and A (mod p) can be transformed to E (mod p) via a series of elementary row operations modulo p. (4) Exercise 7.4. (a) The echelon form (over Q) is   1 0 −1  0 1 2 . 0 0 0

199

11.8. Chapter 8

(b) The kernel is the 1-dimensional span of (1, −2, 1). (c) The characteristic polynomial is x · (x2 − 15x − 18).

(5) Exercise 7.5. (a) The answer is given in the problem. (b) See [Coh93, §2.4].

11.8. Chapter 8 (1) Exercise 8.1. Using the Chinese Remainder Theorem we immediately reduce to proving the statement when both M = pr and N = ps are powers of a prime p. Then [a] ∈ (Z/ps Z)∗ is represented by an integer a with gcd(a, p) = 1. That same integer a defines an element of (Z/pr Z)∗ that reduces modulo ps to [a]. (2) Exercise 8.2. See [Shi94, Lemma 1.38]. (3) Exercise 8.3. Coset representatives for Γ1 (3) are in bijection with (c, d) where c, d ∈ Z/3Z and gcd(c, d, N ) = 1, so the following are representatives: ( 10 01 ) , ( 20 02 ) , ( 01 20 ) , ( 11 01 ) , ( 21 02 ) , ( 12 10 ) , ( 12 01 ) , ( 22 02 ) , which we call r1 , . . . , r8 , respectively. Now our Manin symbols are of the form [X, ri ] and [Y, ri ] for 1 ≤ i ≤ 8 modulo the relations x + xσ = 0,

x + xτ + xτ 2 = 0,

and x − xJ = 0.

First, note that J acts trivially on Manin symbols of odd weight because it sends X to −X, Y to −Y and ri to −ri , so [z, g]J = [−z, −g] = [z, g]. Thus the last relation is trivially true. Now σ −1 X = −Y and σ −1 Y = X. Also τ −1 X = −Y, τ −1 Y = X − Y, τ −2 X = −X + Y and τ −2 Y = −X. The first relation on the first symbol says that [X, r1 ] = −[−Y, r3 ] = [Y, r3 ] and the second relation tells us that [X, r1 ] + [−Y, r5 ] + [−X + Y, r6 ] = 0. (4) Exercise 8.4. Let f ∈ Sk (Γ) and g ∈ Γ. All that remains to be shown is that this pairing respects the relation x = xg for all modular symbols x. By linearity it suffices to show the invariance

200

11. Solutions to Selected Exercises

of hf, X k−i−2 Y i {α, β}i. We have D E f, (X k−2−i Y i {α, β})g −1 D E = f, (aX + bY )k−i−2 (cX + dY )i {g −1 (α), g −1 (β)} Z g−1 (β) f (z)(az + b)k−i−2 (cz + d)i dz = g −1 (α)

=

Z

g −1 (β)

f (z)

g −1 (α)

=

Z

g −1 (β)

(az + b)k−i−2 (cz + d)k−2 dz (cz + d)k−i−2

f (z) g(z)k−i−2 (cz + d)k−2 dz

g −1 (α)

= = =

Z

β

α Z β

α Z β

f (g −1 (z)) g(g −1 (z))k−i−2 (cg −1 (z) + d)k−2 d(g −1 (z)) f (g −1 (z)) z k−i−2 (cg −1 (z) + d)k−2 (cg −1 (z) + d)2 dz f (z) z k−i−2 dz

α

D E = f, X k−i−2 Y i {α, β} ,

where the second to last simplification is due to invariance under [g]k , i.e., f (g −1 (z)) = f [g]k (g −1 (z)) = (cg −1 (z) + d)−k f (g(g −1 (z))). (The proof for f ∈ S k (Γ) works in exactly the same way.)

(5) Exercise 8.5. ³ (a) Let η = ³ −a γη = −c

−1 0 b d

³ ´ a b . For any γ = c d we have ´ ³ ´ ³ −a −b a , ηγ = , and ηγη = c d −c 0 1

´

First, if γ ∈ SL2 (Z), then ηγη ∈ GL2 (Z) and

−b d

´

.

det(ηγη) = det η det γ det η = (−1)(1)(−1) = 1 so ηγη ∈ SL2 (Z). As η 2 = 1, conjugation by η is self-inverse, so it must be a bijection. Now if γ ∈ Γ0 (N ), then c ≡ 0 (mod N ), so −c ≡ 0 (mod N ), and so ηγη ∈ Γ0 (N ). Thus ηΓ0 (N )η = Γ0 (N ). If γ ∈ Γ1 (N ), then −c ≡ 0 (mod N ) as before and also a ≡ d ≡ 1 (mod N ), so ηγη ∈ Γ1 (N ). Thus ηΓ1 (N )η = Γ1 (N ). (b) Omitted.

201

11.10. Chapter 10

11.9. Chapter 9 (1) Exercise 9.1. Consider the surjective homomorphism r : SL2 (Z) → SL2 (Z/N Z). Notice that Γ1 (N ) is the exact inverse image of the subgroup H of matrices of the form ( 10 1∗ ) and Γ0 (N ) is the inverse image of the subgroup T of upper triangular matrices. It thus suffices to observe that H is normal in T , which is clear. Finally, the quotient T /H is isomorphic to the group of diagonal matrices in SL2 (Z/N Z)∗ , which is isomorphic to (Z/N Z)∗ . (2) Exercise 9.2. It is enough to show hpi ∈ Z[. . . , Tn , . . .] for primes p, since each hdi can be written in terms of the hpi. Since p ∤ N , we have that Tp2 = Tp2 − hpipk−1 , so

hpipk−1 = Tp2 − Tp2 .

By Dirichlet’s theorem on primes in arithmetic progression, there is a prime q 6= p congruent to p mod N . Since pk−1 and q k−1 are relatively prime, there exist integers a and b such that apk−1 + bq k−1 = 1. Then hpi = hpi(apk−1 + bq k−1 ) = a(Tp 2 − Tp2 ) + b(Tq 2 − Tq2 ) ∈ Z[. . . , Tn , . . .]. (3) Exercise 9.3. Take N = 33. The space S2 (Γ0 (33)) is a direct sum of the two old subspaces coming from S2 (Γ0 (11)) and the new subspace, which has dimension 1. If f is a basis for S2 (Γ0 (11)) and g is a basis for S2 (Γ0 (33))new , then α1 (f ), α3 (f ), g is a basis for S2 (Γ0 (33)) on which all Hecke operators Tn , with gcd(n, 33) = 1, have diagonal matrix. However, the operator T3 on S2 (Γ0 (33)) does not act as a scalar on α1 (f ), so it cannot be in the ring generated by all operators Tn with gcd(n, 33) = 1. (4) Exercise 9.4. Omitted.

11.10. Chapter 10 (1) Exercise 10.1. Hint: Use either repeated integration by parts or a change of variables that relates the integral to the Γ function. (2) Exercise 10.2. See [Cre97a, §2.8].

(3) Exercise 10.3.√ P (a) Let f = −1 q n . Then d = 2, but the nontrivial conjugate of f is −f , so Vf has dimension 1.

202

11. Solutions to Selected Exercises

(b) Choose α ∈ K such that K = Q(α). Write (11.10.1)

f=

d−1 X

αi gi

i=0

with gi ∈ Q[[q]]. Let Wg be the Q-span of the gi , and let Wf = Vf ∩ Q[[q]]. By considering the Gal(Q/Q) conjugates of (11.10.1), we see that the Galois conjugates of f are in the C-span of the gi , so (11.10.2)

d = dimC Vf ≤ dimQ Wg .

Likewise, taking the above modulo O(q n ) for any n, we obtain a matrix equation F = AG, where the columns of F are the Gal(Q/Q)-conjugates of f , the matrix A is the Vandermonde matrix corresponding to α (and its Gal(Q/Q) conjugates), and G has columns gi . Since A is a Vandermonde matrix, it is invertible, so A−1 F = G. Taking the limit as n goes to infinity, we see that each gi is a linear combination of the fi , hence an element of Vf . Thus Wg ⊂ Wf , so (11.10.2) implies that dimQ Wf ≥ d. But Wf ⊗Q C ⊂ Vf so finally d ≤ dimQ Wf = dimC (Wf ⊗Q C) ≤ dimC Vf = d.

(4) Exercise 10.4. See the appendix to Chapter II in [Cre97a], where this example is worked out in complete detail.

Appendix A

Computing in Higher Rank

by Paul E. Gunnells A.1. Introduction This book has addressed the theoretical and practical problems of performing computations with modular forms. Modular forms are the simplest examples of the general theory of automorphic forms attached to a reductive algebraic group G with an arithmetic subgroup Γ; they are the case G = SL2 (R) with Γ a congruence subgroup of SL2 (Z). For such pairs (G, Γ) the Langlands philosophy asserts that there should be deep connections between automorphic forms and arithmetic, connections that are revealed through the action of the Hecke operators on spaces of automorphic forms. There have been many profound advances in recent years in our understanding of these phenomena, for example: • the establishment of the modularity of elliptic curves defined over Q [Wil95, TW95, Dia96, CDT99, BCDT01], • the proof by Harris–Taylor of the local Langlands correspondence [HT01], and • Lafforgue’s proof of the global Langlands correspondence for function fields [Laf02]. Nevertheless, we are still far from seeing that the links between automorphic forms and arithmetic hold in the broad scope in which they are generally 203

204

A. Computing in Higher Rank

believed. Hence one has the natural problem of studying spaces of automorphic forms computationally. The goal of this appendix is to describe some computational techniques for automorphic forms. We focus on the case G = SLn (R) and Γ ⊂ SLn (Z), since the automorphic forms that arise are one natural generalization of modular forms, and since this is the setting for which we have the most tools available. In fact, we do not work directly with automorphic forms, but rather with the cohomology of the arithmetic group Γ with certain coefficient modules. This is the most natural generalization of the tools developed in previous chapters. Here is a brief overview of the contents. Section A.2 gives background on automorphic forms and the cohomology of arithmetic groups and explains why the two are related. In Section A.3 we describe the basic topological tools used to compute the cohomology of Γ explicitly. Section A.4 defines the Hecke operators, describes the generalization of the modular symbols from Chapter 8 to higher rank, and explains how to compute the action of the Hecke operators on the top degree cohomology group. Section A.5 discusses computation of the Hecke action on cohomology groups below the top degree. Finally, Section A.6 briefly discusses some related material and presents some open problems. A.1.1. The theory of automorphic forms is notorious for the difficulty of its prerequisites. Even if one is only interested in the cohomology of arithmetic groups—a small part of the full theory—one needs considerable background in algebraic groups, algebraic topology, and representation theory. This is somewhat reflected in our presentation, which falls far short of being selfcontained. Indeed, a complete account would require a long book of its own. We have chosen to sketch the foundational material and to provide many pointers to the literature; good general references are [BW00, Harb, LS90, Vog97]. We hope that the energetic reader will follow the references and fill many gaps on his/her own. The choice of topics presented here is heavily influenced (as usual) by the author’s interests and expertise. There are many computational topics in the cohomology of arithmetic groups we have completely omitted, including the trace formula in its many incarnations [GP05], the explicit Jacquet– Langlands correspondence [Dem04, SW05], and moduli space techniques [FvdG, vdG]. We encourage the reader to investigate these extremely interesting and useful techniques. A.1.2. Acknowledgements. I thank Avner Ash, John Cremona, Mark McConnell, and Dan Yasaki for helpful comments. I also thank the NSF for support.

A.2. Automorphic Forms and Arithmetic Groups

205

A.2. Automorphic Forms and Arithmetic Groups A.2.1. Let Γ = Γ0 (N ) ⊂ SL2 (Z) be the usual Hecke congruence subgroup of matrices upper-triangular mod N . Let Y0 (N ) be the modular curve Γ\h, and let X0 (N ) be its canonical compactification obtained by adjoining cusps. For any integer k ≥ 2, let Sk (N ) be the space of weight k holomorphic cuspidal modular forms on Γ. According to Eichler–Shimura [Shi94, Chapter 8], we have the isomorphism (A.2.1)

∼

H 1 (X0 (N ); C) − → S2 (N ) ⊕ S2 (N ),

where the bar denotes complex conjugation and where the isomorphism is one of Hecke modules. More generally, for any integer n ≥ 0, let Mn ⊂ C[x, y] be the subspace of degree n homogeneous polynomials. The space Mn admits a representation of Γ by the “change of variables” map ¶ µ a b · p(x, y) = p(dx − by, −cx + ay). (A.2.2) c d fn on the curve X0 (N ).1 Then the analogue of This induces a local system M (A.2.1) for higher-weight modular forms is the isomorphism (A.2.3)

∼ fk−2 ) − H 1 (X0 (N ); M → Sk (N ) ⊕ Sk (N ).

Note that (A.2.3) reduces to (A.2.1) when k = 2.

Similar considerations apply if we work with the open curve Y0 (N ) instead, except that Eisenstein series also contribute to the cohomology. More precisely, let Ek (N ) be the space of weight k Eisenstein series on Γ0 (N ). Then (A.2.3) becomes (A.2.4)

∼ fk−2 ) − H 1 (Y0 (N ); M → Sk (N ) ⊕ Sk (N ) ⊕ Ek (N ).

These isomorphisms lie at the heart of the modular symbols method. A.2.2. The first step on the path to general automorphic forms is a reinterpretation of modular forms in terms of functions on SL2 (R). Let Γ ⊂ SL2 (Z) be a congruence subgroup. A weight k modular form on Γ is a holomorphic function f : h → C satisfying the transformation property ¶ µ a b k ∈ Γ, z ∈ h. f ((az + b)/(cz + d)) = j(γ, z) f (z), γ = c d 1The classic references for cohomology with local systems are [Ste99a, Section 31] and [Eil47, ˇ Ch. V]. A more recent exposition (in the language of Cech cohomology and locally constant sheaves) can be found in [BT82, II.13]. For an exposition tailored to our needs, see [Harb, Section 2.9].

206

A. Computing in Higher Rank

Here j(γ, z) is the automorphy factor cz + d. There are some additional conditions f must satisfy at the cusps of h, but these are not so important for our discussion. The group G = SL2 (R) acts transitively on h, with the subgroup K = SO(2) fixing i. Thus h can be written as the quotient G/K. From this, we see that f can be viewed as a function G → C that is K-invariant on the right and that satisfies a certain symmetry condition with respect to the Γ-action on the left. Of course not every f with these properties is a modular form: some extra data is needed to take the role of holomorphicity and to handle the behavior at the cusps. Again, this can be ignored right now. We can turn this interpretation around as follows. Suppose ϕ is a function G → C that is Γ-invariant on the left, that is, ϕ(γg) = ϕ(g) for all γ ∈ Γ. Hence ϕ can be thought of as a function ϕ : Γ\G → C. We further suppose that ϕ satisfies a certain symmetry condition with respect to the K-action on the right. In particular, any matrix m ∈ K can be written ¶ µ cos θ − sin θ , θ ∈ R, (A.2.5) m= sin θ cos θ with θ uniquely determined modulo 2π. Let ζm be the complex number eiθ . Then the K-symmetry we require is −k ϕ(gm) = ζm ϕ(z),

where k is some fixed nonnegative integer.

m ∈ K,

It turns out that such functions ϕ are very closely related to modular forms: any f ∈ Sk (Γ) uniquely determines such a function ϕf : Γ\G → C. The correspondence is very simple. Given a weight k modular form f , define (A.2.6)

ϕf (g) := f (g · i)j(g, i)−k .

We claim ϕf is left Γ-invariant and satisfies the desired K-symmetry on the right. Indeed, since j satisfies the cocycle property we have

j(gh, z) = j(g, h · z)j(h, z),

ϕf (γg) = f ((γg)·i)j(γg, i)−k = j(γ, g ·i)k f (g ·i)j(γ, g ·i)−k j(g, i)−k = ϕf (g).

Moreover, any m ∈ K stabilizes i. Hence

ϕf (gm) = f ((gm) · i)j(gm, i)−k = f (g · i)j(m, i)−k j(g, m · i)−k .

−k , and thus From (A.2.5) we have j(m, i)−k = (cos θ + i sin θ)−k = ζm −k ϕf (gm) = ζm ϕf (g).

Hence in (A.2.6) the weight and the automorphy factor “untwist” the Γ-action to make ϕf left Γ-invariant. The upshot is that we can study

A.2. Automorphic Forms and Arithmetic Groups

207

modular forms by studying the spaces of functions that arise through the construction (A.2.6). Of course, not every ϕ : Γ\G → C will arise as ϕf for some f ∈ SK (Γ): after all, f is holomorphic and satisfies rather stringent growth conditions. Pinning down all the requirements is somewhat technical and is (mostly) done in the sequel. A.2.3. Before we define automorphic forms, we need to find the correct generalizations of our groups SL2 (R) and Γ0 (N ). The correct setup is rather technical, but this really reflects the power of the general theory, which handles so many different situations (e.g., Maass forms, Hilbert modular forms, Siegel modular forms, etc.). Let G be a connected Lie group, and let K ⊂ G be a maximal compact subgroup. We assume that G is the set of real points of a connected semisimple algebraic group G defined over Q. These conditions mean the following [PR94, §2.1.1]: (1) The group G has the structure of an affine algebraic variety given by an ideal I in the ring R = C[xij , D−1 ], where the variables {xij | 1 ≤ i, j ≤ n} should be interpreted as the entries of an “indeterminate matrix,” and D is the polynomial det(xij ). Both the group multiplication G × G → G and inversion G → G are required to be morphisms of algebraic varieties. The ring R is the coordinate ring of the algebraic group GLn . Hence this condition means that G can be essentially viewed as a subgroup of GLn (C) defined by polynomial equations in the matrix entries of the latter. (2) Defined over Q means that I is generated by polynomials with rational coefficients. (3) Connected means that G is connected as an algebraic variety. (4) Set of real points means that G is the set of real solutions to the equations determined by I. We write G = G(R). (5) Semisimple means that the maximal connected solvable normal subgroup of G is trivial. Example A.1. The most important example for our purposes is the split form of SLn . For this choice we have G = SLn (R) and K = SO(n). Example A.2. Let F/Q be a number field. Then there is a Q-group G such that G(Q) = SLn (F ). The group G is constructed as RF/Q (SLn ), where RF/Q denotes the restriction of scalars from F to Q [PR94, §2.1.2]. For

208

A. Computing in Higher Rank

example, if F is totally real, the group RF/Q (SL2 ) appears when one studies Hilbert modular forms. Let (r, s) be the signature of the field F , so that F ⊗ R ≃ Rr × Cs . Then G = SLn (R)r × SLn (C)s and K = SO(n)r × SU(n)s . Example A.3. Another important example is the split symplectic group Sp2n . This is the group that arises when one studies Siegel modular forms. The group of real points Sp2n (R) is the subgroup of SL2n (R) preserving a fixed nondegenerate alternating bilinear form on R2n . We have K = U(n). A.2.4. To generalize Γ0 (N ), we need the notion of an arithmetic group. This is a discrete subgroup Γ of the group of rational points G(Q) that is commensurable with the set of integral points G(Z). Here commensurable simply means that Γ ∩ G(Z) is a finite index subgroup of both Γ and G(Z); in particular G(Z) itself is an arithmetic group. Example A.4. For the split form of SLn we have G(Z) = SLn (Z) ⊂ G(Q) = SLn (Q). A trivial way to obtain other arithmetic groups is by conjugation: if g ∈ SLn (Q), then g · SLn (Z) · g −1 is also arithmetic.

A more interesting collection of examples is given by the congruence subgroups. The principal congruence subgroup Γ(N ) is the group of matrices congruent to the identity modulo N for some fixed integer N ≥ 1. A congruence subgroup is a group containing Γ(N ) for some N .

In higher dimensions there are many candidates to generalize the Hecke subgroup Γ0 (N ). For example, one can take the subgroup of SLn (Z) that is upper-triangular mod N . From a computational perspective, this choice is not so good since its index in SLn (Z) is large. A better choice, and the one that usually appears in the literature, is to define Γ0 (N ) to be the subgroup of SLn (Z) with bottom row congruent to (0, . . . , 0, ∗) mod N . A.2.5. We are almost ready to define automorphic forms. Let g be the Lie algebra of G, and let U (g) be its universal enveloping algebra over C. Geometrically, g is just the tangent space at the identity of the smooth manifold G. The algebra U (g) is a certain complex associative algebra canonically built from g. The usual definition would lead us a bit far afield, so we will settle for an equivalent characterization: U (g) can be realized as a certain subalgebra of the ring of differential operators on C ∞ (G), the space of smooth functions on G. In particular, G acts on C ∞ (G) by left translations: given g ∈ G and f ∈ C ∞ (G), we define Lg (f )(x) := f (g −1 x). Then U (g) can be identified with the ring of all differential operators on C ∞ (G) that are invariant under left translation. For our purposes the most

A.2. Automorphic Forms and Arithmetic Groups

209

important part of U (g) is its center Z(g). In terms of differential operators, Z(g) consists of those operators that are also invariant under right translation: Rg (f )(x) := f (xg). Definition A.5. An automorphic form on G with respect to Γ is a function ϕ : G → C satisfying (1) ϕ(γg) = ϕ(g) for all γ ∈ Γ,

(2) the right translates {ϕ(gk) | k ∈ K} span a finite-dimensional space ξ of functions, (3) there exists an ideal J ⊂ Z(g) of finite codimension such that J annihilates ϕ, and (4) ϕ satisfies a certain growth condition that we do not wish to make precise. (In the literature, ϕ is said to be slowly increasing.) For fixed ξ and J, we denote by A (Γ, ξ, J, K) the space of all functions satisfying the above four conditions. It is a basic theorem, due to HarishChandra [HC68], that A (Γ, ξ, J, K) is finite-dimensional. Example A.6. We can identify the cuspidal modular forms Sk (N ) in the language of Definition A.5. Given a modular form f , let ϕf ∈ C ∞ (SL2 (R)) be the function from (A.2.6). Then the map f 7→ ϕf identifies Sk (N ) with the subspace Ak (N ) of functions ϕ satisfying (1) ϕ(γg) = ϕ(g) for all γ ∈ Γ0 (N ),

−k ϕ(g) for all m ∈ SO(2), (2) ϕ(gm) = ζm

(3) (∆ + λk )ϕ = 0, where ∆ ∈ Z(g) is the Laplace–Beltrami–Casimir operator and ¶ µ k k −1 , λk = 2 2 (4) ϕ is slowly increasing, and (5) ϕ is cuspidal. The first four conditions parallel Definition A.5. Item (1) is the Γinvariance. Item (2) implies that the right translates of ϕ by SO(2) lie in a fixed finite-dimensional representation of SO(2). Item (3) is how holomorphicity appears, namely that ϕ is killed by a certain differential operator. Finally, item (4) is the usual growth condition. The only condition missing from the general definition is (5), which is an extra constraint placed on ϕ to ensure that it comes from a cusp form. This condition can be expressed by the vanishing of certain integrals (“constant terms”); for details we refer to [Bum97, Gel75].

210

A. Computing in Higher Rank

Example A.7. Another important example appears when we set k = 0 in (2) in Example A.6 and relax (3) by requiring only that (∆ − λ)ϕ = 0 for some nonzero λ ∈ R. Such automorphic forms cannot possibly arise from modular forms, since there are no nontrivial cusp forms of weight 0. However, there are plenty of solutions to these conditions: they correspond to real-analytic cuspidal modular forms of weight 0 and are known as Maass forms. Traditionally one writes λ = (1 − s2 )/4. The positivity of ∆ implies that s ∈ (−1, 1) or is purely imaginary.

Maass forms are highly elusive objects. Selberg proved that there are infinitely many linearly independent Maass forms of full level (i.e., on SL2 (Z)), but to this date no explicit construction of a single one is known. (Selberg’s argument is indirect and relies on the trace formula; for an exposition see [Sar03].) For higher levels some explicit examples can be constructed using theta series attached to indefinite quadratic forms [Vig77]. Numerically Maass forms have been well studied; see for example [FL].

In general the arithmetic nature of the eigenvalues λ that correspond to Maass forms is unknown, although a famous conjecture of Selberg states that for congruence subgroups they satisfy the inequality λ ≥ 1/4 (in other words, only purely imaginary s appear above). The truth of this conjecture would have far-reaching consequences, from analytic number theory to graph theory [Lub94]. A.2.6. As Example A.6 indicates, there is a notion of cuspidal automorphic form. The exact definition is too technical to state here, but it involves an appropriate generalization of the notion of constant term familiar from modular forms. There are also Eisenstein series [Lan66, Art79]. Again the complete definition is technical; we only mention that there are different types of Eisenstein series corresponding to certain subgroups of G. The Eisenstein series that are easiest to understand are those built from cusp forms on lower rank groups. Very explicit formulas for Eisenstein series on GL3 can be seen in [Bum84]. For a down-to-earth exposition of some of the Eisenstein series on GLn , we refer to [Gol05]. The decomposition of Mk (Γ0 (N )) into cusp forms and Eisenstein series also generalizes to a general group G, although the statement is much more complicated. The result is a theorem of Langlands [Lan76] known as the spectral decomposition of L2 (Γ\G). A thorough recent presentation of this can be found in [MW94]. A.2.7. Let A = A (Γ, K) be the space of all automorphic forms, where ξ and J range over all possibilities. The space A is huge, and the arithmetic significance of much of it is unknown. This is already apparent for G =

A.2. Automorphic Forms and Arithmetic Groups

211

SL2 (R). The automorphic forms directly connected with arithmetic are the holomorphic modular forms, not the Maass forms2. Thus the question arises: which automorphic forms in A are the most natural generalization of the modular forms? One answer is provided by the isomorphisms (A.2.1), (A.2.3), (A.2.4). These show that modular forms appear naturally in the cohomology of modular curves. Hence a reasonable approach is to generalize the left of (A.2.1), (A.2.3), (A.2.4), and to study the resulting cohomology groups. This is the approach we will take. One drawback is that it is not obvious that our generalization has anything to do with automorphic forms, but we will see eventually that it certainly does. So we begin by looking for an appropriate generalization of the modular curve Y0 (N ). Let G and K be as in Section A.2.3, and let X be the quotient G/K. This is a global Riemannian symmetric space [Hel01]. One can prove that X is contractible. Any arithmetic group Γ ⊂ G acts on X properly discontinuously. In particular, if Γ is torsion-free, then the quotient Γ\X is a smooth manifold. Unlike the modular curves, Γ\X will not have a complex structure in general3; nevertheless, Γ\X is a very nice space. In particular, if Γ is torsionfree, it is an Eilenberg–Mac Lane space for Γ, otherwise known as a K(Γ, 1). This means that the only nontrivial homotopy group of Γ\X is its fundamental group, which is isomorphic to Γ, and that the universal cover of Γ\X is contractible. Hence Γ\X is in some sense a “topological incarnation”4 of Γ. This leads us to the notion of the group cohomology H ∗ (Γ; C) of Γ with trivial complex coefficients. In the early days of algebraic topology, this was defined to be the complex cohomology of an Eilenberg–Mac Lane space for Γ [Bro94, Introduction, I.4]: (A.2.7)

H ∗ (Γ; C) = H ∗ (Γ\X; C).

Today there are purely algebraic approaches to H ∗ (Γ; C) [Bro94, III.1], but for our purposes (A.2.7) is exactly what we need. In fact, the group cohomology H ∗ (Γ; C) can be identified with the cohomology of the quotient Γ\X even if Γ has torsion, since we are working with complex coefficients. The cohomology groups H ∗ (Γ; C), where Γ is an arithmetic group, are our proposed generalization for the weight 2 modular forms. What about higher weights? For this we must replace the trivial coefficient module C with local systems, just as we did in (A.2.3). For our 2However, Maass forms play a very important indirect role in arithmetic. 3The symmetric spaces that have a complex structure are known as bounded domains, or Hermitian symmetric spaces [Hel01]. 4This apt phrase is due to Vogan [Vog97].

212

A. Computing in Higher Rank

purposes it is enough to let M be a rational finite-dimensional representation of G over the complex numbers. Any such M gives a representation of f on Γ\X. As before, the group Γ ⊂ G and thus induces a local system M f). In (A.2.3) we took cohomology H ∗ (Γ; M ) is the cohomology H ∗ (Γ\X; M M = Mn , the nth symmetric power of the standard representation. For a general group G there are many kinds of representations to consider. In any case, we contend that the cohomology spaces f) H ∗ (Γ; M ) = H ∗ (Γ\X; M

are a good generalization of the spaces of modular forms. A.2.8. It is certainly not obvious that the cohomology groups H ∗ (Γ; M ) have anything to do with automorphic forms, although the isomorphisms (A.2.1), (A.2.3), (A.2.4) look promising. The connection is provided by a deep theorem of Franke [Fra98], which asserts that (1) the cohomology groups H ∗ (Γ; M ) can be directly computed in terms of certain automorphic forms (the automorphic forms of “cohomological type,” also known as those with “nonvanishing (g, K) cohomology” [VZ84]); and (2) there is a direct sum decomposition M ∗ ∗ H{P (A.2.8) H ∗ (Γ; M ) = Hcusp (Γ; M ) ⊕ } (Γ; M ), {P }

where the sum is taken over the set of classes of associate proper Q-parabolic subgroups of G. The precise version of statement (1) is known in the literature as the Borel conjecture. Statement (2) parallels Langlands’s spectral decomposition of L2 (Γ\G). Example A.8. For Γ = Γ0 (N ) ⊂ SL2 (Z), the decomposition (A.2.8) is exactly (A.2.4). The cusp forms Sk (N )⊕Sk (N ) correspond to the summand 1 (Γ; M ). There is one class of proper Q-parabolic subgroups in SL (R), Hcusp 2 represented by the Borel subgroup of upper-triangular matrices. Hence only one term appears in big direct sum on the right of (A.2.8), which is the Eisenstein term Ek . ∗ (Γ; M ) of (A.2.8) is called the cuspidal cohomolThe summand Hcusp ogy; this is the subspace of classes represented by cuspidal automorphic forms. The remaining summands constitute the Eisenstein cohomology of Γ [Har91]. In particular the summand indexed by {P } is constructed using Eisenstein series attached to certain cuspidal automorphic forms on lower

A.3. Combinatorial Models for Group Cohomology

213

∗ (Γ; M ) is in some sense the most important part rank groups. Hence Hcusp of the cohomology: all the rest can be built systematically from cuspidal cohomology on lower rank groups5. This leads us to our basic computational problem:

Problem A.9. Develop tools to compute explicitly the cohomology spaces ∗ (Γ; M ). H ∗ (Γ; M ) and to identify the cuspidal subspace Hcusp

A.3. Combinatorial Models for Group Cohomology A.3.1. In this section, we restrict attention to G = SLn (R) and Γ, a congruence subgroup of SLn (Z). By the previous section, we can study the f). group cohomology H ∗ (Γ; M ) by studying the cohomology H ∗ (Γ\X; M The latter spaces can be studied using standard topological techniques, such as taking the cohomology of complexes associated to cellular decompositions of Γ\X. For SLn (R), one can construct such decompositions using a version of explicit reduction theory of real positive-definite quadratic forms due to Voronoˇı [Vor08]. The goal of this section is to explain how this is done. We also discuss how the cohomology can be explicitly studied for congruence subgroups of SL3 (Z). A.3.2. Let V be the R-vector space of all symmetric n × n matrices, and let C ⊂ V be the subset of positive-definite matrices. The space C can be identified with the space of all real positive-definite quadratic forms in n variables: in coordinates, if x = (x1 , . . . , xn )t ∈ Rn (column vector), then the matrix A ∈ C induces the quadratic form x 7−→ xt Ax,

and it is well known that any positive-definite quadratic form arises in this way. The space C is a cone, in that it is preserved by homotheties: if x ∈ C, then λx ∈ C for all λ ∈ R>0 . It is also convex: if x1 , x2 ∈ C, then tx1 + (1 − t)x2 ∈ C for t ∈ [0, 1]. Let D be the quotient of C by homotheties.

Example A.10. The case n = 2 is illustrative. We can take coordinates on V ≃ R3 by representing any matrix in V as ¶ µ x y , x, y, z ∈ R. y z The subset of singular matrices Q = {xz − y 2 = 0} is a quadric cone in V dividing the complement V r Q into three connected components. The component containing the identity matrix is the cone C of positive-definite matrices. The quotient D can be identified with an open 2-disk. 5This is a bit of an oversimplification, since it is a highly nontrivial problem to decide when cusp cohomology from lower rank groups appears in Γ. However, many results are known; as a selection we mention [Har91, Har87, LS04]

214

A. Computing in Higher Rank

The group G acts on C on the left by (g, c) 7−→ gcg t .

This action commutes with that of the homotheties and thus descends to a G-action on D. One can show that G acts transitively on D and that the stabilizer of the image of the identity matrix is K = SO(n). Hence we may identify D with our symmetric space X = SLn (R)/ SO(n). We will do this in the sequel, using the notation D when we want to emphasize the coordinates coming from the linear structure of C ⊂ V and using the notation X for the quotient G/K. We can make the identification D ≃ X more explicit. If g ∈ SLn (R), then the map (A.3.1)

g 7−→ gg t

takes g to a symmetric positive-definite matrix. Any coset gK is taken to the same matrix since KK t = Id. Thus (A.3.1) identifies G/K with a subset C1 of C, namely those positive-definite symmetric matrices with determinant 1. It is easy to see that C1 maps diffeomorphically onto D. The inverse map C1 → G/K is more complicated. Given a determinant 1 positive-definite symmetric matrix A, one must find g ∈ SLn (R) such that gg t = A. Such a representation always exists, with g determined uniquely up to right multiplication by an element of K. In computational linear algebra, such a g can be constructed through Cholesky decomposition of A. The group SLn (Z) acts on C via the G-action and does so properly discontinuously. This is the “unimodular change of variables” action on quadratic forms [Ser73, V.1.1]. Under our identification of D with X, this is the usual action of SLn (Z) by left translation from Section A.2.7. f). A.3.3. Now consider the group cohomology H ∗ (Γ; M ) = H ∗ (Γ\X; M The identification D ≃ X shows that the dimension of X is n(n + 1)/2 − 1. Hence H i (Γ; M ) vanishes if i > n(n + 1)/2 − 1. Since dim X grows quadratically in n, there are many potentially interesting cohomology groups to study. However, it turns out that there is some additional vanishing of the cohomology for deeper (topological) reasons. For n = 2, this is easy to see. The quotient Γ\h is homeomorphic to a topological surface with punctures, corresponding to the cusps of Γ. Any such surface S can be retracted onto a finite graph simply by “stretching” S along its punctures. Thus H 2 (Γ; M ) = 0, even though dim Γ\h = 2. For Γ ⊂ SLn (Z), a theorem of Borel–Serre implies that H i (Γ; M ) vanishes if i > dim X − n + 1 = n(n − 1)/2 [BS73, Theorem 11.4.4]. The number ν = n(n − 1)/2 is called the virtual cohomological dimension of Γ

215

A.3. Combinatorial Models for Group Cohomology

and is denoted vcd Γ. Thus we only need to consider cohomology in degrees i ≤ ν.

Moreover we know from Section A.2.8 that the most interesting part of the cohomology is the cuspidal cohomology. In what degrees can it live? For n = 2, there is only one interesting cohomology group H 1 (Γ; M ), and it contains the cuspidal cohomology. For higher dimensions, the situation i is quite different: for most i, the subspace Hcusp (Γ; M ) vanishes! In fact in the late 1970’s Borel, Wallach, and Zuckerman observed that the cuspidal cohomology can only live in the cohomological degrees lying in an interval around (dim X)/2 of size linear in n. An explicit description of this interval is given in [Sch86, Proposition 3.5]; one can also look at Table A.3.1, from which the precise statement is easy to determine. Another feature of Table A.3.1 deserves to be mentioned. There are exactly two values of n, namely n = 2, 3, such that virtual cohomological dimension equals the upper limit of the cuspidal range. This will have implications later, when we study the action of the Hecke operators on the cohomology. n dim X vcd Γ ∗ top degree of Hcusp ∗ bottom degree of Hcusp

2 2 1 1 1

3 5 3 3 2

4 5 6 7 8 9 14 20 27 35 6 10 15 21 28 5 8 11 15 19 4 6 9 12 16

9 44 36 24 20

Table A.3.1. The virtual cohomological dimension and the cuspidal range for subgroups of SLn (Z).

A.3.4. Recall that a point in Zn is said to be primitive if the greatest common divisor of its coordinates is 1. In particular, a primitive point is nonzero. Let P ⊂ Zn be the set of primitive points. Any v ∈ P, written as a column vector, determines a rank-1 symmetric matrix q(v) in the closure C¯ via q(v) = vv t . The Voronoˇı polyhedron Π is defined to be the closed convex hull in C¯ of the points q(v), as v ranges over P. Note that by construction, SLn (Z) acts on Π, since SLn (Z) preserves the set {q(v)} and acts linearly on V . Example A.11. Figure A.3.1 represents a crude attempt to show what Π looks like for n = 2. These images were constructed by computing a large subset of the points q(v) and taking the convex hull (we took all points v ∈ P such that Trace q(v) < N for some large integer N ). From a distance, the polyhedron Π looks almost indistinguishable from the cone C; this is somewhat conveyed by the right of Figure A.3.1. Unfortunately Π is not

216

A. Computing in Higher Rank

locally finite, so we really cannot produce an accurate picture. To get a more accurate image, the reader should imagine that each vertex meets infinitely many edges. On the other hand, Π is not hopelessly complex: each maximal face is a triangle, as the pictures suggest.

(a)

(b)

Figure A.3.1. The polyhedron Π for SL2 (Z). In (a) we see Π from the origin, in (b) from the side. The small triangle at the right center of (a) is the facet with vertices {q(e1 ), q(e2 ), q(e1 + e2 )}, where {e1 , e2 } is the standard basis of Z2 . In (b) the x-axis runs along the top from left to right, and the z-axis runs down the left side. The facet from (a) is the little triangle at the top left corner of (b).

A.3.5. The polyhedron Π is quite complicated: it has infinitely many faces and is not locally finite. However, one of Voronoˇı’s great insights is that Π is actually not as complicated as it seems. For any A ∈ C, let µ(A) be the minimum value attained by A on P and let M (A) ⊂ P be the set on which A attains µ(A). Note that µ(A) > 0 and M (A) is finite since A is positive-definite. Then A is called perfect if it is recoverable from the knowledge of the pair (µ(A), M (A)). In other words, given (µ(A), M (A)), we can write a system of linear equations (A.3.2)

mZmt = µ(A),

m ∈ M (A),

where Z = (zij ) is a symmetric matrix of variables. Then A is perfect if and only if A is the unique solution to the system (A.3.2). Example A.12. The quadratic form Q(x, y) = x2 − xy + y 2 is perfect. The smallest nontrivial value it attains on Z2 is µ(Q) = 1, and it does so on the columns of ¶ µ 1 0 1 M (Q) = 0 1 1

A.3. Combinatorial Models for Group Cohomology

217

and their negatives. Letting αx2 + βxy + γy 2 be an undetermined quadratic form and applying the data (µ(Q), M (Q)), we are led to the system of linear equations α = 1, γ = 1, α + β + γ = 1. From this we recover Q(x, y). Example A.13. The quadratic form Q′ (x, y) = x2 + y 2 is not perfect. Again the smallest nontrivial value of Q′ on Z2 is m(Q′ ) = 1, attained on the columns of ¶ µ 1 0 ′ M (Q ) = 0 1 and their negatives. But every member of the one-parameter family of quadratic forms (A.3.3)

x2 + αxy + y 2 ,

α ∈ (−1, 1)

has the same set of minimal vectors, and so Q′ cannot be recovered from the knowledge of m(Q′ ), M (Q′ ). Example A.14. Example A.12 generalizes to all n. Define (A.3.4)

An (x) :=

n X i=1

x2i −

X

xi xj .

1≤i 2 and let Γ = SLn (Z). The picture is very similar, except that now there are several Hecke operators attached to any prime p. In fact there are n − 1 operators T (p, k), k = 1, . . . , n − 1. The operator T (p, k) is associated to the correspondence C(g), where g = diag(1, . . . , 1, p, . . . , p) and where p occurs k times. If we consider the congruence subgroups Γ0 (N ), we have operators T (p, k) for (p, N ) = 1 and analogues of the Up operators for p|N . Just as in the classical case, any double coset ΓgΓ can be written as a disjoint union of left cosets a Γh ΓgΓ = h∈Ω

8Under the identification H ∗ (Γ\X; M f) ≃ H ∗ (Γ; M ), the map t∗ becomes the transfer map in group cohomology [Bro94, III.9].

A.4. Hecke Operators and Modular Symbols

227

for a certain finite set of n × n integral matrices Ω. For the operator T (p, k), the set Ω can be taken to be all upper-triangular matrices of the form [Kri90, Proposition 7.2]  e  p1 aij   ..  , . e n p where • ei ∈ {0, 1} and exactly k of the ei are equal to 1 and

• aij = 0 unless ei = 0 and ej = 1, in which case aij satisfies 0 ≤ aij < p. Remark A.19. The number of coset representatives for the operator T (p, k) is the same as the number of points in the finite Grassmannian G(k, n)(Fp ). A similar phenomenon is true for the Hecke operators for any group G, although there are some subtleties [Gro98]. A.4.3. Recall that in Section A.3.6 we constructed the Voronoˇı decomposition C and the well-rounded retract W and that we can use them to compute the cohomology H ∗ (Γ; M ). Unfortunately, we cannot directly use them to compute the action of the Hecke operators on cohomology, since the Hecke operators do not act cellularly on C or W . The problem is that the Hecke image of a cell in C (or W ) is usually not a union of cells in C (or W ). This is already apparent for n = 2. The edges of C are the SL2 (Z)-translates of the ideal geodesic τ from 0 to ∞ (Example A.15). Applying a Hecke operator takes such an edge to a union of ideal geodesics, each with vertices at a pair of cusps. In general such geodesics are not an SL2 (Z)-translate of τ . For n = 2, one solution is to work with all possible ideal geodesics with vertices at the cusps, in other words the space of modular symbols M2 from Section 3.2. Manin’s trick (Proposition 3.11) shows how to write any modular symbol as a linear combination of unimodular symbols, by which we mean modular symbols supported on the edges of C . These are the ideas we now generalize to all n. Definition A.20. Let S0 be the Q-vector space spanned by the symbols v = [v1 , . . . , vn ], where vi ∈ Qn r {0}, modulo the following relations: (1) If τ is a permutation on n letters, then

[v1 , . . . , vn ] = sign(τ )[τ (v1 ), . . . , τ (vn )], where sign(τ ) is the sign of τ . (2) If q ∈ Q× , then

[qv1 , v2 . . . , vn ] = [v1 , . . . , vn ].

228

A. Computing in Higher Rank

(3) If the points v1 , . . . , vn are linearly dependent, then v = 0. Let B ⊂ S0 be the subspace generated by linear combinations of the form (A.4.1)

n X

(−1)i [v0 , . . . , vˆi , . . . , vn ],

i=0

where v0 , . . . , vn ∈

Qn

r {0} and where vˆi means to omit vi .

We call S0 the space of modular symbols. We caution the reader that there are some differences in what we call modular symbols and those found in Section 3.2 and Definition 8.2; we compare them in Section A.4.4. The group SLn (Q) acts on S0 by left multiplication: g · v = [gv1 , . . . , gvn ]. This action preserves the subspace B and thus induces an action on the quotient M = S0 /B. For Γ ⊂ SLn (Z) a finite index subgroup, let MΓ be the space of Γ-coinvariants in M . In other words, MΓ is the quotient of M by the subspace generated by {m − γ · m | γ ∈ Γ}. The relationship between modular symbols and the cohomology of Γ is given by the following theorem, first proved for SLn by Ash and Rudolph [AR79] and by Ash for general G [Ash86]:

Theorem A.21 ([Ash86, AR79]). Let Γ ⊂ SLn (Z) be a finite index subgroup. There is an isomorphism (A.4.2)

∼

MΓ − → H ν (Γ; Q),

where Γ acts trivially on Q and where ν = vcd(Γ). We remark that Theorem A.21 remains true if Q is replaced with nontrivial coefficients as in Section A.2.7. Moreover, if Γ is assumed to be torsion-free then we can replace Q with Z. The great virtue of MΓ is that it admits an action of the Hecke operators. Given a Hecke operator Tg , write the double coset ΓgΓ as a disjoint union of left cosets a Γh (A.4.3) ΓgΓ = h∈Ω

as inPExample A.18. Any class in MΓ can be lifted to a representative η = q(v)v ∈ S0 , where q(v) ∈ Q and almost all q(v) vanish. Then we define X (A.4.4) Tg (v) = h·v h∈Ω

and extend to η by linearity. The right side of (A.4.4) depends on the choices of η and Ω, but after taking quotients and coinvariants, we obtain a well-defined action on cohomology via (A.4.2).

A.4. Hecke Operators and Modular Symbols

229

A.4.4. The space S0 is closely related to the space M2 from Section 3.2 and Section 8.1. Indeed, M2 was defined to be the quotient (F/R)/(F/R)tor , where F is the free abelian group generated by ordered pairs (A.4.5)

{α, β},

α, β ∈ P1 (Q),

and R is the subgroup generated by elements of the form (A.4.6)

{α, β} + {β, γ} + {γ, α},

α, β, γ ∈ P1 (Q).

The only new feature in Definition A.20 is item (3). For n = 2 this corresponds to the condition {α, α} = 0, which follows from (A.4.6). We have S0 /B ≃ M2 ⊗ Q. Hence there are two differences between S0 and M2 : our notion of modular symbols uses rational coefficients instead of integral coefficients and is the space of symbols before dividing out by the subspace of relations B; we further caution the reader that this is somewhat at odds with the literature. We also remark that the general arbitrary weight definition of modular symbols for a subgroup Γ ⊂ SL2 (Z) given in Section 8.1 also includes taking Γ-coinvariants, as well as extra data for a coefficient system. We have not included the latter data since our emphasis is trivial coefficients, although it would be easy to do so in the spirit of Section 8.1. Elements of M2 also have a geometric interpretation: the symbol {α, β} corresponds to the ideal geodesic in h with endpoints at the cusps α and β. We have a similar picture for the symbols v = [v1 , . . . , vn ]. We can assume that each vi is primitive, which means that each vi determines a vertex of the Voronoˇı polyhedron Π. The rational cone generated by these vertices determines a subset ∆(v) ⊂ D, where D is the linear model of the symmetric space X = SLn (R)/ SO(n) from Section A.3.2. This subset ∆(v) is then an “ideal simplex” in X. There is also a connection between ∆(v) and torus orbits in X; we refer to [Ash86] for a related discussion. A.4.5. Now we need a generalization of the Manin trick (Section 3.3.1). This is known in the literature as the modular symbols algorithm. We can define a kind of norm function on S0 as follows. Let v = [v1 , . . . , vn ] be a modular symbol. For each vi , choose λi ∈ Q× such that λi vi is primitive. Then we define kvk := | det(λ1 v1 , . . . , λn vn )| ∈ Z. Note that kvk is well defined, since the λi are unique up to sign, and permuting the vi only changes the determinant by a sign. We extend k k to all of S0 by taking the maximum of k k over the support of any η ∈ S0 : if

230

η=

A. Computing in Higher Rank P

q(v)v, where q(v) ∈ Q and almost all q(v) vanish, then we put kηk = Max kvk. q(v)6=0

We say a modular symbol η is unimodular if kηk = 1. It is clear that the images of the unimodular symbols generate a finite-dimensional subspace of MΓ . The next theorem shows that this subspace is actually all of MΓ . Theorem A.22 ([AR79, Bar94]). The space MΓ is spanned by the images of the unimodular symbols. More precisely, given any symbol v ∈ S0 with kvk > 1, (1) in S0 /B we may write X (A.4.7) v= q(w)w,

q(w) ∈ Z,

where if q(w) 6= 0, then kwk = 1, and

(2) the number of terms on the right side of (A.4.7) is bounded by a polynomial in log kvk that depends only on the dimension n. Proof. (Sketch) Given a modular symbol v = [v1 , . . . , vn ], we may assume that the points vi are primitive. We will show that if kvk > 1, we can find a point u such that when we apply the relation (A.4.1) using the points u, v1 , . . . , vn , all terms other than v have norm less than kvk. We call such a point a reducing point for v. Let P ⊂ Rn be the open parallelotope ¯ nX o ¯ P := λi vi ¯ |λi | < kvk−1/n .

Then P is an n-dimensional centrally symmetric convex body with volume 2n . By Minkowski’s theorem from the geometry of numbers (cf. [FT93, IV.2.6]), P ∩ Zn contains a nonzero point u. Using (A.4.1), we find (A.4.8)

v=

n X

(−1)i−1 vi (u),

i=1

where vi (u) is the symbol

vi (u) = [v1 , . . . , vi−1 , u, vi+1 , . . . , vn ]. Moreover, it is easy to see that the new symbols satisfy (A.4.9)

0 ≤ kvi (u)k < kvk(n−1)/n ,

i = 1, . . . , n.

This completes the proof of the first statement. To prove the second statement, we must estimate how many times relations of the form (A.4.8) need to be applied to obtain (A.4.7). A nonunimodular symbol produces at most n new modular symbols after (A.4.8) is

A.4. Hecke Operators and Modular Symbols

231

performed; we potentially have to apply (A.4.8) again to each of the symbols that result, which in turn could produce as many as n new symbols for each. Hence we can visualize the process of constructing (A.4.7) as building a rooted tree, where the root is v, the leaves are the symbols w, and where each node has at most n children. It is not hard to see that the bound (A.4.9) implies that the depth of this tree (i.e., the longest length of a path from the root to a leaf) is O(log log kvk). From this the second statement follows easily. ¤ Statement (1) of Theorem A.22 is due to Ash and Rudolph [AR79]. Instead of P , they used the larger parallelotope P ′ defined by ¯ o nX ¯ P ′ := λi vi ¯ |λi | < 1 , which has volume 2n kvk. The observation that P ′ can be replaced by P and the proof of (2) are both due to Barvinok [Bar94].

A.4.6. The relationship between Theorem A.22 and Manin’s trick should be clear. For Γ ⊂ SL2 (Z), the Manin symbols correspond exactly to the unimodular symbols mod Γ. So Theorem A.22 implies that every modular symbol (in the language of Section 8.1) is a linear combination of Manin symbols. This is exactly the conclusion of Proposition 8.3. In higher rank the relationship between Manin symbols and unimodular symbols is more subtle. In fact there are two possible notions of “Manin symbol,” which agree for SL2 (Z) but not in general. One possibility is the obvious one: a Manin symbol is a unimodular symbol. The other possibility is to define a Manin symbol to be a modular symbol corresponding to a top-dimensional cell of the retract W . But for n ≥ 5, such modular symbols need not be unimodular. In particular, for n = 5 there are two equivalence classes of top-dimensional cells. One class corresponds to the unimodular symbols, the other to a set of modular symbols of norm 2. However, Theorems A.21 and A.22 show that H ν (Γ; Q) is spanned by unimodular symbols. Thus as far as this cohomology group is concerned, the second class of symbols is in some sense unnecessary. A.4.7. We return to the setting of Section A.3.8 and give examples of Hecke eigenclasses in the cusp cohomology of Γ = Γ0 (p) ⊂ SL3 (Z). We closely follow [AGG84, vGvdKTV97]. Note that since the top of the cuspidal range for SL3 is the same as the virtual cohomological dimension ν, we can use modular symbols to compute the Hecke action on cuspidal classes. Given a prime l coprime to p, there are two Hecke operators of interest T (l, 1) and T (l, 2). We can compute the action of these operators on

232

A. Computing in Higher Rank

3 (Γ; C) as follows. Recall that H 3 (Γ; C) can be identified with a cerHcusp cusp tain space of functions f : P2 (Fp ) → C (Theorem A.16). Given x ∈ P2 (Fp ), let Qx ∈ SL3 (Z) be a matrix such that Qx 7→ x under the identification ∼ P2 (Fp ) − → Γ\ SL3 (Z). Then Qx determines a unimodular symbol [Qx ] by taking the vi to be the columns of Qx . Given ` any Hecke operator Tg , we can find coset representatives hi such that ΓgΓ = Γhi (explicit representatives for Γ = Γ0 (p) and Tg = T (l, k) are given in [AGG84, vGvdKTV97]). The modular symbols [hi Qx ] are no longer unimodular in general, but we can apply Theorem A.22 to write X [hi Qx ] = [Rij ], Rij ∈ SL3 (Z). j

Then for f : P2 (Fp ) → C as in Theorem A.16, we have X (Tg f )(x) = f (Rij ), i,j

where Rij is the class of Rij in

P2 (F

p ).

3 (Γ; C) Hcusp

Now let ξ ∈ be a simultaneous eigenclass for all the Hecke operators T (l, 1), T (l, 2), as l ranges over all primes coprime with p. General considerations from the theory of automorphic forms imply that the eigenvalues a(l, 1), a(l, 2) are complex conjugates of one other. Hence it suffices to compute a(l, 1). We give two examples of cuspidal eigenclasses for two different prime levels. 3 (Γ (53); C) is 2-dimensional. Let Example A.23. Let p = 53. Then Hcusp 0 √ η = (1 + −11)/2. One eigenclass is given by the data

l a(l, 1)

2 −1 − 2η

3 −2 + 2η

5 1

7 −3

11 1

and the other is obtained by complex conjugation.

13 −2 − 12η

3 (Γ (61); C) is 2-dimensional. Let Example A.24. Let p = 61. Then Hcusp 0 √ ω = (1 + −3)/2. One eigenclass is given by the data

l a(l, 1)

2 1 − 2ω

3 −5 + 4ω

5 −2 + 4ω

7 −6ω

and the other is obtained by complex conjugation.

11 −2 + 2ω

13 −2 − 4ω

A.5. Other Cohomology Groups A.5.1. In Section A.4 we saw how to compute the Hecke action on the top cohomology group H ν (Γ; C). Unfortunately for n ≥ 4, this cohomology group does not contain any cuspidal cohomology. The first case is Γ ⊂ SL4 (Z); we have vcd(Γ) = 6, and the cusp cohomology lives in degrees 4

233

A.5. Other Cohomology Groups

and 5. One can show that the cusp cohomology in degree 4 is dual to that in degree 5, so for computational purposes it suffices to be able to compute the Hecke action on H 5 (Γ; C). But modular symbols do not help us here. In this section we describe a technique to compute the Hecke action on H ν−1 (Γ; C), following [Gun00a]. The technique is an extension of the modular symbol algorithm to these cohomology groups. In principle the ideas in this section can be modified to compute the Hecke action on other cohomology groups H ν−k (Γ; C), k > 1, although this has not been investigated9. For n = 4, we have applied the algorithm in joint work with Ash and McConnell to investigate computationally the cohomology H 5 (Γ; C), where Γ0 (N ) ⊂ SL4 (Z) [AGM02]. A.5.2. To begin, we need an analogue of Theorem A.21 for lower degree cohomology groups. In other words, we need a generalization of the modular symbols for other cohomology groups. This is achieved by the sharbly complex S∗ : Definition A.25 ([Ash94]). Let {S∗ , ∂} be the chain complex given by the following data: (1) For k ≥ 0, Sk is the Q-vector space generated by the symbols u = [v1 , . . . , vn+k ], where vi ∈ Qn r {0}, modulo the relations: (a) If τ is a permutation on (n + k) letters, then [v1 , . . . , vn+k ] = sign(τ )[τ (v1 ), . . . , τ (vn+k )], where sign(τ ) is the sign of τ . (b) If q ∈ Q× , then [qv1 , v2 . . . , vn+k ] = [v1 , . . . , vn+k ]. (c) If the rank of the matrix (v1 , . . . , vn+k ) is less than n, then u = 0. (2) For k > 0, the boundary map ∂ : Sk → Sk−1 is [v1 , . . . , vn+k ] 7−→

n+k X

(−1)i [v1 , . . . , vˆi , . . . , vn+k ].

i=1

We define ∂ to be identically zero on S0 . The elements u = [v1 , . . . , vn+k ] 9The first interesting case is n = 5, for which the cuspidal cohomology lives in H ν−2 .

234

A. Computing in Higher Rank

are called k-sharblies10. The 0-sharblies are exactly the modular symbols from Definition A.20, and the subspace B ⊂ S0 is the image of the boundary map ∂ : S1 → S0 . There is an obvious left action of Γ on S∗ commuting with ∂. For any k ≥ 0, let Sk,Γ be the space of Γ-coinvariants. Since the boundary map ∂ commutes with the Γ-action, we obtain a complex (S∗,Γ , ∂Γ ). The following theorem shows that this complex computes the cohomology of Γ: Theorem A.26 ([Ash94]). There is a natural isomorphism ∼

H ν−k (Γ; C) − → Hk (S∗,Γ ⊗ C). A.5.3. We can extend our norm function k k from modular symbols to all of Sk as follows. Let u = [v1 , . . . , vn+k ] be a k-sharbly, and let Z(u) be the set of all submodular symbols determined by u. In other words, Z(u) consists of the modular symbols of the form [vi1 , . . . , vin ], where {i1 , . . . , in } ranges over all n-fold subsets of {1, . . . , n + k}. Define kuk by kuk = Max kvk. v∈Z(u)

Note that kuk is well defined modulo the relations in Definition A.25. As P for modular symbols, we extend the norm to sharbly chains ξ = q(u)u taking the maximum normSover the support. Formally, we let supp(ξ) = {u | q(u) 6= 0} and Z(ξ) = u∈supp(ξ) Z(u), and then we define kξk by kξk = Max kvk. v∈Z(ξ)

We say that ξ is reduced if kξk = 1. Hence ξ is reduced if and only if all its submodular symbols are unimodular or have determinant 0. Clearly there are only finitely many reduced k-sharblies modulo Γ for any k. In general the cohomology groups H ∗ (Γ; C) are not spanned by reduced sharblies. However, it is known (cf. [McC91]) that for Γ ⊂ SL4 (Z), the group H 5 (Γ; C) is spanned by reduced 1-sharbly cycles. The best one can say in general is that for each pair n, k, there is an integer N = N (n, k) such that for Γ ⊂ SLn (Z), H ν−k (Γ; C) is spanned by k-sharblies of norm ≤ N . This set of sharblies is also finite modulo Γ, although it is not known how large N must be for any given pair n, k. A.5.4. Recall that the cells of the well-rounded retract W are indexed by sets of primitive vectors in Zn . Since any primitive vector determines a point in Qn r{0} and since sets of such points index sharblies, it is clear that there is a close relationship between S∗ and the chain complex associated to W , 10The terminology for S is due to Lee Rudolph, in honor of Lee and Szczarba. They ∗ introduced a very similar complex in [LS76] for SL3 (Z).

235

A.5. Other Cohomology Groups

although of course S∗ is much bigger. In any case, both complexes compute H ∗ (Γ; C). The main benefit of using the sharbly complex P to compute cohomology is that it admits a Hecke action. Suppose ξ = q(u)u is a sharbly cycle mod Γ, and consider a Hecke operator Tg . Then we have X n(u)h · u, (A.5.1) Tg (ξ) = h∈Ω,u

where Ω is a set of coset representatives as in (A.4.3). Since Ω 6⊂ SLn (Z) in general, the Hecke image of a reduced sharbly is not usually reduced. A.5.5. We are now ready to describe our algorithm for the computation of the Hecke operators on H ν−1 (Γ; C). It suffices to describe an algorithm that takes as input a 1-sharbly cycle ξ and produces as output a cycle ξ ′ with (a) the classes of ξ and ξ ′ in H ν−1 (Γ; C) the same, and (b) kξ ′ k < kξk if kξk > 1. Below, we will present an algorithm satisfying (a). In [Gun00a], we conjectured (and presented evidence) that the algorithm satisfies (b) for n ≤ 4. Further evidence is provided by the computations in [AGM02], which relied on the algorithm to compute the Hecke action on H 5 (Γ; C), where Γ = Γ0 (N ) ⊂ SL4 (Z).

The idea behind the algorithm is simple: given a 1-sharbly cycle ξ that is not reduced, (i) simultaneously apply the modular symbol algorithm (Theorem A.22) to each of its submodular symbols, and then (ii) package the resulting data into a new 1-sharbly cycle. Our experience in presenting this algorithm is that most people find the geometry involved in (ii) daunting. Hence we will give details only for n = 2 and will provide a sketch for n > 2. Full details are contained in [Gun00a]. Note that n = 2 is topologically and arithmetically uninteresting, since we are computing the Hecke action on H 0 (Γ; C); nevertheless, the geometry faithfully represents the situation for all n. A.5.6. Fix n = 2, let ξ ∈ S1 be a 1-sharbly cycle mod Γ for some Γ ⊂ SL2 (Z), and suppose ξ is not reduced. Assume Γ is torsion-free to simplify the presentation. Suppose first that all submodular symbols v ∈ Z(ξ) are nonunimodular. Select reducing points for each v ∈ Z(ξ) and make these choices Γ-equivariantly. This means the following. Suppose u, u′ ∈ supp ξ and v ∈ supp(∂u) and v′ ∈ supp(∂u′ ) are modular symbols such that v = γ · v′ for some γ ∈ Γ. Then we select reducing points w for v and w′ for v′ such

236

A. Computing in Higher Rank

that w = γ · w′ . (Note that since Γ is torsion-free, no modular symbol can be identified to itself by an element of Γ; hence v 6= v′ .) This is possible since if v is a modular symbol and w is a reducing point for v, then γ · w is a reducing point for γ · v for any γ ∈ Γ. Because there are only finitely many Γ-orbits in Z(ξ), we can choose reducing points Γ-equivariantly by selecting them for some set of orbit representatives. It is important to note that Γ-equivariance is the only global criterion we use when selecting reducing. In particular, there is a priori no relationship among the three reducing points chosen for any u ∈ supp ξ. A.5.7. Now we want to use the reducing points and the 1-sharblies in ξ to build ξ ′ . Choose u = [v1 , v2 , v3 ] ∈ supp ξ, and denote the reducing point for [vi , vj ] by wk , where {i, j, k} = {1, 2, 3}. We use the vi and the wi to build a 2-sharbly chain η(u) as follows. Let P be an octahedron in R3 . Label the vertices of P with the vi and wi such that the vertex labeled vi is opposite the vertex labeled wi (Figure A.5.1). Subdivide P into four tetrahedra by connecting two opposite vertices, say v1 and w1 , with an edge (Figure A.5.2). For each tetrahedron T , take the labels of four vertices and arrange them into a quadruple. If we orient P , then we can use the induced orientation on T to order the four primitive points. In this way, each T determines a 2-sharbly, and η(u) is defined to be the sum. For example, if we use the decomposition in Figure A.5.2, we have (A.5.2) η(u) = [v1 , v3 , v2 , w1 ] + [v1 , w2 , v3 , w1 ] + [v1 , w3 , w2 , w1 ] + [v1 , v2 , w3 , w1 ]. P Repeat this construction for all u ∈ supp ξ, and let η = q(u)η(u). Finally, let ξ ′ = ξ + ∂η. w1 w2 w3 v2

v3 v1

Figure A.5.1.

A.5.8. By construction, ξ ′ is a cycle mod Γ in the same class as ξ. We claim in addition that no submodular symbol from ξ appears in ξ ′ . To see this, consider ∂η(u). From (A.5.2), we have (A.5.3) ∂η(u) = −[v1 , v2 , v3 ] + [v1 , v2 , w3 ] + [v1 , w2 , v3 ] + [w1 , v2 , v3 ]

− [v1 , w2 , w3 ] − [w1 , v2 , w3 ] − [w1 , w2 , v3 ] + [w1 , w2 , w3 ].

A.5. Other Cohomology Groups

237

Figure A.5.2.

Note that this is the boundary in S∗ , not in S∗,Γ . Furthermore, ∂η(u) is independent of which pair of opposite vertices of P we connected to build η(u). From (A.5.3), we see that in ξ + ∂η the 1-sharbly −[v1 , v2 , v3 ] is canceled by u ∈ supp ξ. We also claim that 1-sharblies in (A.5.3) of the form [vi , vj , wk ] vanish in ∂Γ η. To see this, let u,u′ ∈ supp ξ, and suppose v = [v1 , v2 ] ∈ supp ∂u equals γ · v′ for some v′ = [v1′ , v2′ ] ∈ supp ∂u′ . Since the reducing points were chosen Γ-equivariantly, we have w = γ · w′ . This means that the 1-sharbly [v1 , v2 , w] ∈ ∂η(u) will be canceled mod Γ by [v1′ , v2′ , w′ ] ∈ ∂η(u′ ). Hence, in passing from ξ to ξ ′ , the effect in (S∗ )Γ is to replace u with four 1-sharblies in supp ξ ′ : (A.5.4) [v1 , v2 , v3 ] 7−→ −[v1 , w2 , w3 ]−[w1 , v2 , w3 ]−[w1 , w2 , v3 ]+[w1 , w2 , w3 ]. Note that in (A.5.4), there are no 1-sharblies of the form [vi , vj , wk ]. Remark A.27. For implementation purposes, it is not necessary to explicitly construct η. Rather, one may work directly with (A.5.4). A.5.9. Why do we expect ξ ′ to satisfy kξ ′ k < kξk? First of all, in the right hand side of (A.5.4) there are no submodular symbols of the form [vi , vj ]. In fact, any submodular symbol involving a point vi also includes a reducing point for [vi , vj ]. On the other hand, consider the submodular symbols in (A.5.4) of the form [wi , wj ]. Since there is no relationship among the wi , one has no reason to believe that these modular symbols are closer to unimodularity than those in u. Indeed, for certain choices of reducing points it can happen that k[wi , wj ]k ≥ kuk.

The upshot is that some care must be taken in choosing reducing points. In [Gun00a, Conjectures 3.5 and 3.6] we describe two methods for finding reducing points for modular symbols, one using Voronoˇı reduction and one using LLL-reduction. Our experience is that if one selects reducing points using either of these conjectures, then k[wi , wj ]k < kuk for each of the new modular symbols [wi , wj ]. In fact, in practice these symbols are trivial or satisfy k[wi , wj ]k = 1.

238

A. Computing in Higher Rank

A.5.10. In the previous discussion we assumed that no submodular symbols of any u ∈ supp ξ were unimodular. Now we say what to do if some are. There are three cases to consider. First, all submodular symbols of u may be unimodular. In this case there are no reducing points, and (A.5.4) becomes (A.5.5)

[v1 , v2 , v3 ] 7−→ [v1 , v2 , v3 ].

Second, one submodular symbol of u may be nonunimodular, say the symbol [v1 , v2 ]. In this case, to build η, we use a tetrahedron P ′ and put η(u) = [v1 , v2 , v3 , w3 ] (Figure A.5.3). Since [v1 , v2 , w3 ] vanishes in the boundary of η mod Γ, (A.5.4) becomes (A.5.6)

[v1 , v2 , v3 ] 7→ −[v1 , v3 , w3 ] + [v2 , v3 , w3 ]. v1 w3 v3

v2

Figure A.5.3.

Finally, two submodular symbols of u may be nonunimodular, say [v1 , v2 ] and [v1 , v3 ]. In this case we use the cone on a square P ′′ (Figure A.5.4). To construct η(u), we must choose a decomposition of P ′′ into tetrahedra. Since P ′′ has a nonsimplicial face, this choice affects ξ ′ (in contrast to the previous cases). If we subdivide P ′′ by connecting the vertex labelled v2 with the vertex labelled w2 , we obtain (A.5.7)

[v1 , v2 , v3 ] 7−→ [v2 , w2 , w3 ] + [v2 , v3 , w2 ] + [v1 , v3 , w2 ]. v1 w2

w3

v3

v2

Figure A.5.4.

A.5.11. Now consider general n. The basic technique is the same, but the combinatorics become more complicated. Suppose u = [v1 , . . . , vn+1 ] satisfies q(u) 6= 0 in a 1-sharbly cycle ξ, and for i = 1, . . . , n + 1 let vi be the submodular symbol [v1 , . . . , vbi , . . . , vn+1 ]. Assume that all vi are nonunimodular, and for each i let wi be a reducing point for vi .

239

A.5. Other Cohomology Groups

For any subset I ⊂ {1, . . . , n + 1}, let uI be the 1-sharbly [u1 , . . . , un+1 ], where ui = wi if i ∈ I, and ui = vi otherwise. The polytope P used to build η(u) is the cross polytope, which is the higher-dimensional analogue of the octahedron [Gun00a, §4.4]. We suppress the details and give the final answer: (A.5.4) becomes X (A.5.8) u 7−→ − (−1)#I uI , I

where the sum is taken over all subsets I ⊂ {1, . . . , n + 1} of cardinality at least 2.

More generally, if some vi happen to be unimodular, then the polytope used to build η is an iterated cone on a lower-dimensional cross polytope. This is already visible for n = 2: • The 2-dimensional cross polytope is a square, and the polytope P ′′ is a cone on a square. • The 1-dimensional cross polytope is an interval, and the polytope P ′ is a double cone on an interval. Altogether there are n + 1 relations generalizing (A.5.5)–(A.5.7). A.5.12. Now we describe how these computations are carried out in practice, focusing on Γ = Γ0 (N ) ⊂ SL4 (Z) and H 5 (Γ; C). Besides discussing technical details, we also have to slightly modify some aspects of the construction in Section A.5.6, since Γ is not torsion-free. Let W be the well-rounded retract. We can represent a cohomology class P 5 β ∈ H (Γ; C) as β = q(σ)σ, where σ denotes a codimension 1 cell in W . In this case there are three types of codimension 1 cells in W . Under the bijection W ↔ C , these cells correspond to the Voronoˇı cells indexed by the columns of the matrices (A.5.9)       1 0 0 0 1 1 0 0 0 1 1 0 0 0 1      0 1 0 0 1  ,  0 1 0 0 1 ,  0 1 0 0 1 .   0 0 1 0 0   0 0 1 0 1   0 0 1 0 1  0 0 0 1 0 0 0 0 1 0 0 0 0 1 1 Thus each σ in W modulo Γ corresponds to an SL4 (Z)-translate of one of the matrices in (A.5.9). These translates determine basis 1-sharblies u (by taking the points ui toP be the columns), and hence we can represent β by a 1-sharbly chain ξ = q(u)u ∈ S1 that is a cycle in the complex of coinvariants (S∗,Γ , ∂Γ ).

To make later computations more efficient, we precompute more data attached to ξ. Given a 1-sharbly u = [u1 , . . . , un+1 ], a lift M (u) of u is defined to be an integral matrix with primitive columns Mi such that u =

240

A. Computing in Higher Rank

[M1 , . . . , Mn+1 ]. Then we encode ξ, once and for all, by a finite collection Φ of 4-tuples (u, n(u), {v}, {M (v)}), where (1) u ranges over the support of ξ, (2) n(u) ∈ C is the coefficient of u in ξ,

(3) {v} is the set of submodular symbols appearing in the boundary of u, and

(4) {M (v)} is a set of lifts for {v}. Moreover, the lifts in (4) are chosen to satisfy the following Γ-equivariance condition. Suppose that for u, u′ ∈ supp ξ we have v ∈ supp(∂u) and v′ ∈ supp(∂u′ ) satisfying v = γ · v′ for some γ ∈ Γ. Then we require M (v) = γM (v′ ). This is possible since ξ is a cycle modulo Γ, although there is one complication since Γ has torsion: it can happen that some submodular symbol v of a 1-sharbly u is identified to itself by an element of Γ. This means that in constructing {M (v)} for u, we must somehow choose more than one lift for v. To deal with this, let M (v) be any lift of v, and let Γ(v) ⊂ Γ be the stabilizer of v. Then in ξ, we replace q(u)u by X 1 q(u)uγ , #Γ(v) γ∈Γ(v)

where uγ has the same data as u, except11 that we give v the lift γM (v). Next we compute and store the 1-sharbly transformation laws generalizing (A.5.5)–(A.5.7). As a part of this we fix triangulations of certain cross polytopes as in (A.5.7). We are now ready to begin the actual reduction algorithm. We take a Hecke operator T (l, k) and build the coset representatives Ω as in (A.5.1). For each h ∈ Ω and each 1-sharbly u in the support of ξ, we obtain a nonreduced 1-sharbly uh := h · u. Here h acts on all the data attached to u in the list Φ. In particular, we replace each lift M (v) with h · M (v), where the dot means matrix multiplication. Now we check the submodular symbols of uh and choose reducing points for the nonunimodular symbols. This is where the lifts come in handy. Recall that reduction points must be chosen Γ-equivariantly over the entire cycle. Instead of explicitly keeping track of the identifications between modular symbols, we do the following trick: 11In fact, we can be slightly more clever than this and only introduce denominators that are powers of 2.

A.5. Other Cohomology Groups

241

(1) Construct the Hermite normal form Mher (v) of the lift M (v) (see [Coh93, §2.4] and Exercise 7.5). Record the transformation matrix U ∈ GL4 (Z) such that U M (v) = Mher (v). (2) Choose a reducing point u for Mher (v). (3) Then the reducing point for M (v) is U −1 u. This guarantees Γ-equivariance: if v, v′ are submodular symbols of ξ with γ ·v = v′ and with reducing points u, u′ , we have γu = u′ . The reason is that the Hermite normal form Mher (v) is a uniquely determined representative of the GL4 (Z)-orbit of M (v) [Coh93]. Hence if γM (v) = M (v′ ), then Mher (v) = Mher (v′ ). After computing all reducing points, we apply the appropriate transformation law. The result will be a chain of 1-sharblies, each of which has (conjecturally) smaller norm than the original 1-sharbly u. We output these 1-sharblies if they are reduced; otherwise they are fed into the reduction algorithm again. Eventually we obtain a reduced 1-sharbly cycle ξ ′ homologous to the original cycle ξ. The final step of the algorithm is to rewrite ξ ′ as a cocycle on W . This is easy to do since the relevant cells of W are in bijection with the reduced 1-sharblies. There are some nuisances in keeping orientations straight, but the computation is not difficult. We refer to [AGM02] for details.

A.5.13. We now give some examples, taken from [AGM02], of Hecke eigenclasses in H 5 (Γ0 (N ); C) for various levels N . Instead of giving a table of eigenvalues, we give the Hecke polynomials. If β is an eigenclass with T (l, k)(β) = a(l, k)β, then we define H(β, l) =

X (−1)k lk(k−1)/2 a(l, k)X k ∈ C[X]. k

For almost all l, after putting X = l−s where s is a complex variable, the function H(β, s) is the inverse of the local factor at l of the automorphic representation attached to β. Example A.28. Suppose N = 11. Then the cohomology H 5 (Γ0 (11); C) is 2-dimensional. There are two Hecke eigenclasses u1 , u2 , each with rational Hecke eigenvalues.

242

A. Computing in Higher Rank

u1

u2

T2 T3 T5 T7 T2 T3 T5 T7

(1 − 4X)(1 − 8X)(1 + 2X + 2X 2 ) (1 − 9X)(1 − 27X)(1 + X + 3X 2 ) (1 − 25X)(1 − 125X)(1 − X + 5X 2 ) (1 − 49X)(1 − 343X)(1 + 2X + 7X 2 ) (1 − X)(1 − 2X)(1 + 8X + 32X 2 ) (1 − X)(1 − 3X)(1 + 9X + 243X 2 ) (1 − X)(1 − 5X)(1 − 25X + 3125X 2 ) (1 − X)(1 − 7X)(1 + 98X + 16807X 2 )

Example A.29. Suppose N = 19. Then the cohomology H 5 (Γ0 (19); C) is 3-dimensional. There are three Hecke eigenclasses u1 , u2 , u3 , each with rational Hecke eigenvalues. u1

u2

u3

T2 T3 T5 T2 T3 T5 T2 T3 T5

(1 − 4X)(1 − 8X)(1 + 2X 2 ) (1 − 9X)(1 − 27X)(1 + 2X + 3X 2 ) (1 − 25X)(1 − 125X)(1 − 3X + 5X 2 ) (1 − X)(1 − 2X)(1 + 32X 2 ) (1 − X)(1 − 3X)(1 + 18X + 243X 2 ) (1 − X)(1 − 5X)(1 − 75X + 3125X 2 ) (1 − 2X)(1 − 4X)(1 + 3X + 8X 2 ) (1 − 3X)(1 − 9X)(1 + 5X + 27X 2 ) (1 − 5X)(1 − 25X)(1 + 12X + 125X 2 )

In these examples, the cohomology is completely accounted for by the Eisenstein summand of (A.2.8). In fact, let Γ′0 (N ) ⊂ SL2 (Z) be the usual Hecke congruence subgroup of matrices upper-triangular modulo N . Then the cohomology classes above actually come from classes in H 1 (Γ′0 (N )), that is from holomorphic modular forms of level N . For N = 11, the space of weight two cusp forms S2 (11) is 1-dimensional. This cusp form f lifts in two different ways to H 5 (Γ0 (11); C), which can be seen from the quadratic part of the Hecke polynomials for the ui . Indeed, for ui the quadratic part is exactly the inverse of the local factor for the Lfunction attached to f , after the substitution X = l−s . For u2 , we see that the lift is also twisted by the square of the cyclotomic character. (In fact the linear terms of the Hecke polynomials come from powers of the cyclotomic character.) For N = 19, the space of weight two cusp forms S2 (19) is again 1dimensional. The classes u1 and u2 are lifts of this form, exactly as for N = 11. The class u3 , on the other hand, comes from S4 (19), the space of weight 4 cusp forms on Γ′0 (19). In fact, dim S4 (19) = 4, with one Hecke eigenform defined over Q and another defined over a totally real cubic extension of Q. Only the rational weight four eigenform contributes to H 5 (Γ0 (19); C). One can show that whether or not a weight four cuspidal

A.5. Other Cohomology Groups

243

eigenform f contributes to the cohomology of Γ0 (N ) depends only on the sign of the functional equation of L(f, s) [Wes]. This phenomenon is typical of what one encounters when studying Eisenstein cohomology. In addition to the lifts of weight 2 and weight 4 cusp forms, for other levels one finds lifts of Eisenstein series of weights 2 and 4 and lifts of cuspidal cohomology classes from subgroups of SL3 (Z). For some levels one finds cuspidal classes that appear to be lifts from the group of symplectic similitudes GSp(4). More details can be found in [AGM02, AGM]. A.5.14. Here are some notes on the reduction algorithm and its implementation: • Some additional care must be taken when selecting reducing points for the submodular symbols of u. P In particular, in practice one should choose w for v such that kvi (w)k is minimized. Similar remarks apply when choosing a subdivision of the crosspolytopes in Section A.5.10. • In practice, the reduction algorithm has always terminated with a reduced 1-sharbly cycle ξ ′ homologous to ξ. However, at the moment we cannot prove that this will always happen. • Experimentally, the efficiency of the reduction step appears to be comparable to that of Theorem A.22. In other words the depth of the “reduction tree” associated to a given 1-sharbly u seems to be bounded by a polynomial in log log kuk. Hence computing the Hecke action using this algorithm is extremely efficient. On the other hand, computing Hecke operators on SL4 is still a much bigger computation—relative to the level—than on SL2 and SL3 . For example, the size of the full retract W modulo Γ0 (p) is roughly O(p6 ), which grows rapidly with p. The portion of the retract corresponding to H 5 is much smaller, around p3 /10, but this still grows quite quickly. This makes computing with p > 100 out of reach at the moment. The number of Hecke cosets grows rapidly as well, e.g., the number of coset representatives of T (l, 2) is l4 + l3 + 2l2 + l + 1. Hence it is only feasible to compute Hecke operators for small l; for large levels only l = 2 is possible. Here are some numbers to give an idea of the size of these computations. For level 73, the rank of H 5 is 20. There are 39504 cells of codimension 1 and 4128 top-dimensional cells in W modulo Γ0 (73). The computational techniques in [AGM02] used at this level (a Lanczos scheme over a large finite field) tend to produce sharbly cycles supported on almost all the cells. Computing T (2, 1)

244

A. Computing in Higher Rank

requires a reduction tree of depth 1 and produces as many as 26 reduced 1-sharblies for each of the 15 nonreduced Hecke images. Thus one cycle produces a cycle supported on as many as 15406560 1-sharblies, all of which must be converted to an appropriate cell of W modulo Γ. Also this is just what needs to be done for one cycle; do not forget that the rank of H 5 is 20. In practice the numbers are slightly better, since the reduction step produces fewer 1-sharblies on average and since the support of the initial cycle has size less than 39504. Nevertheless the orders of magnitude are correct. • Using lifts is a convenient way to encode the global Γ-identifications in the cycle ξ, since it means we do not have to maintain a big data structure keeping track of the identifications on ∂ξ. However, there is a certain expense in computing the Hermite normal form. This is balanced by the benefit that working with the data Φ associated to ξ allows us to reduce the supporting 1-sharblies u independently. This means we can cheaply parallelize our computation: each 1sharbly, encoded as a 4-tuple (u, n(u), {v}, {M (v)}), can be handled by a separate computer. The results of all these individual computations can then be collated at the end, when producing a W -cocycle.

A.6. Complements and Open Problems A.6.1. We conclude this appendix by giving some complements and describing some possible directions for future work, both theoretical and computational. Since a full explanation of the material in this section would involve many more pages, we will be brief and will provide many references. A.6.2. Perfect Quadratic Forms over Number Fields and Retracts. Since Voronoˇı’s pioneering work [Vor08], it has been the goal of many to extend his results from Q to a general algebraic number field F . Recently Coulangeon [Cou01], building on work of Icaza and Baeza [Ica97, BI97], has found a good notion of perfection for quadratic forms over number fields12. One of the key ideas in [Cou01] is that the correct notion of equivalence between Humbert forms involves not only the action of GLn (OF ), where OF is the ring of integers of F , but also the action of a certain continuous group U related to the units OF× . One of Coulangeon’s basic results is that there are finitely many equivalence classes of perfect Humbert forms modulo these actions. 12Such forms are called Humbert forms in the literature.

A.6. Complements and Open Problems

245

On the other hand, Ash’s original construction of retracts [Ash77] introduces a geometric notion of perfection. Namely he generalizes the Voronoˇı polyhedron Π and defines a quadratic form to be perfect if it naturally indexes a facet of Π. What is the connection between these two notions? Can one use Coulangeon’s results to construct cell complexes to be used in cohomology computations? One tempting possibility is to try to use the group U to collapse the Voronoˇı cells of [Ash77] into a cell decomposition of the symmetric space associated to SLn (F ). A.6.3. The Modular Complex. In his study of multiple ζ-values, Goncharov has recently defined the modular complex M ∗ [Gon97, Gon98]. This is an n-step complex of GLn (Z)-modules closely related both to the properties of multiple polylogarithms evaluated at µN , the N th roots of unity, and to the action of GQ on π1,N = π1l (P1 r {0, ∞, µN }), the pro-l completion of the algebraic fundamental group of P1 r {0, ∞, µN }.

Remarkably, the modular complex is very closely related to the Voronoˇı decomposition V . In fact, one can succinctly describe the modular complex by saying that it is the chain complex of the cells coming from the top-dimensional Voronoˇı cone of type An . This is all of the Voronoˇı decomposition for n = 2, 3, and Goncharov showed that the modular complex is quasi-isomorphic to the full Voronoˇı complex for n = 4. Hence there is a precise relationship among multiple polylogarithms, the Galois action on π1,N , and the cohomology of level N congruence subgroups of SLn (Z).

The question then arises, how much of the cohomology of congruence subgroups is captured by the modular complex for all n? Table A.3.2 indicates that asymptotically very little of the Voronoˇı decomposition comes from the An cone, but this says nothing about the cohomology. The first interesting case to consider is n = 5. A.6.4. Retracts for Other Groups. The most general construction of retracts W known [Ash84] applies only to linear symmetric spaces. The most familiar example of such a space is SLn (R)/ SO(n); other examples are the symmetric spaces associated to SLn over number fields and division algebras. Now let Γ ⊂ G(Q) be an arithmetic group, and let X = G/K be the associated symmetric space. What can one say about cell complexes that can be used to compute H ∗ (Γ; M )? The theorem of Borel–Serre mentioned in Section A.3.3 implies the vanishing of H k (Γ; M ) for k > ν := dim X − q, where q is the Q-rank of Γ. For example, for the split form of SLn , the Q-rank is n − 1. For the split symplectic group Sp2n , the Q-rank is n. Moreover, this bound is sharp: there will be coefficient modules M for

246

A. Computing in Higher Rank

which H ν (Γ; M ) 6= 0. Hence any minimal cell complex used to compute the cohomology of Γ should have dimension ν. Ideally one would like to see such a complex realized as a subspace of X and would like to be able to treat all finite index subgroups of Γ simultaneously. This leads to the following question: is there a Γ-equivariant deformation retraction of X onto a regular cell complex W of dimension ν? For G = Sp4 , McConnell and MacPherson showed that the answer is yes. Their construction begins by realizing the symplectic symmetric space XSp as a subspace of the special linear symmetric space XSL . They then construct subsets of XSp by intersecting the Voronoˇı cells in XSL with XSp . Through explicit computations in coordinates they prove that these intersections are cells and give a cell decomposition of XSp . By taking an appropriate dual complex (as suggested by Figures A.3.2 and A.3.3 and as done in [Ash77]), they construct the desired cell complex W . Other progress has been recently made by Bullock [Bul00], Bullock and Connell [BC06], and Yasaki [Yas05b, Yas05a] in the case of groups of Qrank 1. In particular, Yasaki uses the tilings of Saper [Sap97] to construct an explicit retract for the unitary group SU(2, 1) over the Gaussian integers. His method also works for Hilbert modular groups, although further refinement may be needed to produce a regular cell complex. Can one generalize these techniques to construct retracts for groups of arbitrary Q-rank? Is there an analogue of the Voronoˇı decomposition for these retracts (i.e., a dual cell decomposition of the symmetric space)? If so, can one generalize ideas in Sections A.4–A.5 and use that generalization to compute the action of the Hecke operators on the cohomology? A.6.5. Deeper Cohomology Groups. The algorithm in Section A.5 can be used to compute the Hecke action on H ν−1 (Γ). For n > 4, this group no longer contains cuspidal cohomology classes. Can one generalize this algorithm to compute the Hecke action on deeper cohomology groups? The first practical case is n = 5. Here ν = 10, and the highest degree in which cuspidal cohomology can live is 8. This case is also interesting since the cohomology of full level has been studied [EVGS02]. Here are some indications of what one can expect. The general strategy is the same: for a k-sharbly ξ representing a class in H ν−k (Γ), begin by Γequivariantly choosing reducing points for the nonunimodular submodular symbols of ξ. This data can be packaged into a new k-sharbly cycle as in Section A.5.7ff, but the crosspolytopes must be replaced with hypersimplices. n By Pdefinition, the hypersimplex ∆(n, k) is the convex hull in R of the points { i∈I ei }, where I ranges over all order k subsets of {1, . . . , n} and e1 , . . . , en denotes the standard basis of Rn .

247

A.6. Complements and Open Problems

The simplest example is n = 2, k = 2. From the point of view of cohomology, this is even less interesting than n = 2, k = 1, since now we are computing the Hecke action on H −1 (Γ)! Nevertheless, the geometry here illustrates what one can expect in general. Each 2-sharbly in the support of ξ can be written as [v1 , v2 , v3 , v4 ] and determines six submodular symbols, of the form [vi , vj ], i 6= j. Assume for simplicity that all these submodular symbols are nonunimodular. Let wij be the reducing point for [vi , vj ]. Then use the ten points vi , wij to label the vertices of the hypersimplex ∆(5, 2) as in Figure A.6.1 (note that ∆(5, 2) is 4-dimensional). v1

w14

w12

w13 w24 v2

v4 w34

w23

v3 Figure A.6.1.

The boundary of this hypersimplex gives the analogue of (A.5.4). Which 2-sharblies will appear in ξ ′ ? The boundary ∂∆(5, 2) is a union of five tetrahedra and five octahedra. The outer tetrahedron will not appear in ξ ′ , since that is the analogue of the left side of (A.5.4). The four octahedra sharing a triangular face with the outer tetrahedron also will not appear, since they disappear when considering ξ ′ modulo Γ. The remaining four tetrahedra and the central octahedron survive to ξ ′ and constitute the right side of the analogue of (A.5.4). Note that we must choose a simplicial subdivision of the central octahedron to write the result as a 2-sharbly cycle and that this must be done with care since it introduces a new submodular symbol. If some submodular symbols are unimodular, then again one must consider iterated cones on hypersimplices, just as in Section A.5.10. The analogues of these steps become more complicated, since there are now many

248

A. Computing in Higher Rank

simplicial subdivisions of a hypersimplex13. There is one final complication: in general we cannot use reduced k-sharblies alone to represent cohomology classes. Thus one must terminate the algorithm when kξk is less than some predetermined bound. A.6.6. Other Linear Groups. Let F be a number field, and let G = RF/Q (SLn ) (Example A.2). Let Γ ⊂ G(Q) be an arithmetic subgroup. Can one compute the action of the Hecke operators on H ∗ (Γ)? There are two completely different approaches to this problem. The first involves the generalization of the modular symbols method. One can define the analogue of the sharbly complex, and can try to extend the techniques of Sections A.4–A.5. This technique has been extensively used when F is imaginary quadratic and n = 2. We have X = SL2 (C)/ SU(2), which is isomorphic to 3-dimensional hyperbolic space h3 . The arithmetic groups Γ ⊂ SL2 (OF ) are known as Bianchi groups. The retracts and cohomology of these groups have been well studied; as a representative sample of works we mention [Men79, EGM98, Vog85, GS81]. Such groups have Q-rank 1 and thus have cohomological dimension 2. One can show that the cuspidal classes live in degrees 1 and 2. This means that we can use modular symbols to investigate the Hecke action on cuspidal cohomology. This was done by Cremona [Cre84] for euclidean fields F . In that case Theorem A.22 works with no trouble (the euclidean algorithm is needed to construct reducing points). For noneuclidean fields further work has been done by Whitley [Whi90], Cremona and Whitely √ [CW94] (both for principal ideal domains), Bygott [Byg99] (for F = Q( −5) and any field with class group an elementary abelian 2-group), and Lingham [Lin05] (any field with odd class number). Putting all these ideas together allows one to generalize the modular symbols method to any imaginary quadratic field [Cre]. For F imaginary quadratic and n > 2, very little has been studied. The only related work to the best of our knowledge is that of Staffeldt [Sta79]. He determined the structure of the Voronoˇı polyhedron in detail √ for RF/Q (SL3 ), where F = Q( −1). We have dim X = 8 and ν = 6. The cuspidal cohomology appears in degrees 3, 4, 5, so one could try to use the techniques of Section A.5 to investigate it. Similar remarks apply to F real quadratic and n = 2. The symmetric space X ≃ h × h has dimension 4 and the Q-rank is 1, which means ν = 3. Unfortunately the cuspidal cohomology appears only in degree 2, which 13Indeed, computing all simplicial subdivisions of ∆(n, k) is a difficult problem in convex geometry.

A.6. Complements and Open Problems

249

means modular symbols cannot see it. On the other hand, 1-sharblies can see it, and so one can try to use ideas in Section A.5 here to compute the Hecke operators. The data needed to build the retract W already (essentially) appears in the literature for certain fields; see for example [Ong86]. The second approach shifts the emphasis from modular symbols and the sharbly complex to the Voronoˇı fan and its cones. For this approach we must assume that the group Γ is associated to a self-adjoint homogeneous cone over Q. (cf. [Ash77]). This class of groups includes arithmetic subgroups of RF/Q (SLn ), where F is a totally real or CM field. Such groups have all the nice structures in Section A.3.2. For example, we have a cone C with a G-action. We also have an analogue of the Voronoˇı polyhedron Π. There is a natural compactification C˜ of C obtained by adjoining certain self-adjoint homogeneous cones of lower rank. The quotient Γ\C˜ is singular in general, but it can still be used to compute H ∗ (Γ; C). The polyhedron Π can be used to construct a fan V that gives a Γ-equivariant decomposition of all ˜ But the most important structure we have is the Voronoˇı reduction of C. ˜ we can determine the unique Voronoˇı algorithm: given any point x ∈ C, cone containing x. Here is how this setup can be used to compute the Hecke action. Full details are in [Gun99, GM03]. We define two chain complexes CV∗ and CR ∗ . The latter is essentially the chain complex generated by all simplicial ˜ the former is the subcomplex generated by rational polyhedral cones in C; the Voronoˇı cones. These are the analogues of the sharbly complex and the chain complex associated to the retract W , and one can show that either can be used to compute H ∗ (Γ; C). Take a cycle ξ ∈ CV∗ representing a cohomology class in H ∗ (Γ; C) and act on it by a Hecke operator T . We have V T (ξ) ∈ CR ∗ , and we must push T (ξ) back to C∗ . To do this, we use the linear structure on C˜ to subdivide T (ξ) very finely into a chain ξ ′ . For each 1-cone τ in supp ξ ′ , we choose a 1-cone ρτ ∈ C˜ r C and assemble them using the combinatorics of ξ ′ into a polyhedral chain ξ ′′ homologous to ξ ′ . Under certain conditions involved in the construction of ξ ′ , this chain ξ ′′ will lie in CV∗ .

We illustrate this process for the split group SL2 ; more details can be found in [Gun99]. We work modulo homotheties, so that the threedimensional cone C˜ becomes the extended upper half plane h∗ := h∪Q∪{∞}, with ∂ C˜ passing to the cusps h∗ r h. As usual top-dimensional Voronoˇı cones become the triangles of the Farey tessellation, and the cones ρτ become cusps. Given any x ∈ h, let R(x) be the set of cusps of the unique triangle or edge containing x (this can be computed using the Voronoˇı reduction algorithm). Extend R to a function on h∗ by putting R(u) = {u} for any cusp u.

250

A. Computing in Higher Rank

In h, the support of T (ξ) becomes a geodesic µ between two cusps u, u′ , in other words the support of a modular symbol [u, u′ ] (Figure A.6.2). Subdivide µ by choosing points x0 , . . . , xn such that x0 = u, xn = u′ , and R(xi ) ∩ R(xi+1 ) 6= ∅. (This is easily done, for example by repeatedly barycentrically subdividing µ.) For each i < n choose a cusp qi ∈ R(xi ) ∩ R(xi+1 ), and put qn = u′ . Then we have a relation in H 1 : (A.6.1)

[u, u′ ] = [q0 , q1 ] + · · · + [qn−1 , qn ].

Moreover, each [qi , qi+1 ] is unimodular, since qi and qi+1 are both vertices of a triangle containing xi+1 . Upon lifting (A.6.1) back to CR ∗ , the cusps qi become the 1-cones ρτ and give us a relation T (ξ) = ξ ′′ ∈ CV∗ . µ

u

u′

Figure A.6.2. A subdivision of µ; the solid dots are the xi . Since the xi lie in the same or adjacent Voronoˇı cells, we can assign cusps to them to construct a homology to a cycle in CV∗ .

A.6.7. The Sharbly Complex for General Groups. In [Gun00b] we generalized Theorem A.22 (without the complexity statement) to the symplectic group Sp2n . Using this algorithm and the symplectic retract [MM93, MM89], one can compute the action of the Hecke operators on the topdegree cohomology of subgroups of Sp4 (Z). More recently, Toth has investigated modular symbols for other groups. He showed that the unimodular symbols generate the top-degree cohomology groups for Γ an arithmetic subgroup of a split classical group or a split group of type E6 or E7 [Tot05]. His technique of proof is completely different from that of [Gun00b]. In particular he does not give an analogue of the Manin trick. Can one extract an algorithm from Toth’s proof that can be used to explicitly compute the action of the Hecke operators on cohomology? The proof of the main result of [Gun00b] uses a description of the relations among the modular symbols. These relations were motivated by the structure of the cell complex in [MM93, MM89]. The modular symbols and these relations are analogues of the groups S0 and S1 in the sharbly complex. Can one extend these combinatorial constructions to form a symplectic sharbly complex? What about for general groups G? Already for Sp4 , resolution of this question would have immediate arithmetic applications. Indeed, Harder has a beautiful conjecture about certain

A.6. Complements and Open Problems

251

congruences between holomorphic modular forms and Siegel modular forms of full level [Hara]. Examples of these congruences were checked numerically in [Hara] using techniques of [FvdG] to compute the Hecke action. However, to investigate higher levels, one needs a different technique. The relevant cohomology classes live in H ν−1 (Γ; M ), so one only needs to understand the first three terms of the complex S0 ← S1 ← S2 . We understand S0 , S1 from [Gun00b]; the key is understanding S2 , which should encode relations among elements of S1 . If one could do this and then could generalize the techniques of [Gun00a], one would have a way to investigate Harder’s conjecture.

A.6.8. Generalized Modular Symbols. We conclude this appendix by discussing a geometric approach to modular symbols. This complements the algebraic approaches presented in this book and leads to many new interesting phenomena and problems. Suppose H and G are connected semisimple algebraic groups over Q with an injective map f : H → G. Let KH be a maximal compact subgroup of H = H(R), and suppose K ⊂ G is a maximal compact subgroup containing f (KH ). Let X = G/K and Y = H/KH . Now let Γ ⊂ G(Q) be a torsion-free arithmetic subgroup. Let ΓH = f −1 (Γ). We get a map ΓH \Y → Γ\X, and we denote the image by S(H, Γ). Any compactly supported cohomology class ξ ∈ Hcdim Y (Γ\X; C) can be pulled back via f to ΓH \Y and integrated to obtain a complex number. Hence S(H, Γ) defines a linear form on Hcdim Y (Γ\X; C). By Poincar´e duality, this linear form determines a class [S(H, Γ)] ∈ H dim X−dim Y (Γ\X; C), called a generalized modular symbol. Such classes have been considered by many authors, for example [AB90, SV03, Har05, AGR93]. As an example, we can take G to be the split form of SL2 , and we can take f : H → G to be the inclusion of connected component of the diagonal subgroup. Hence H ≃ R>0 . In this case KH is trivial. The image of Y in X is the ideal geodesic from 0 to ∞. One way to vary f is by taking an SL2 (Q)-translate of this geodesic, which gives a geodesic between two cusps. Hence we can obtain the support of any modular symbol this way. This example generalizes to SLn to yield the modular symbols in Section A.4. Here H ≃ (R > 0)n−1 . Note that dim Y = n − 1, so the cohomology classes we have constructed live in the top degree H ν (Γ\X; C). Another family of examples is provided by taking H to be a Levi factor of a parabolic subgroup; these are the modular symbols studied in [AB90]. There are many natural questions to study for such objects. Here are two:

252

A. Computing in Higher Rank

• Under what conditions on G, H, Γ is [S(H, Γ)] nonzero? This question is connected to relations between periods of automorphic forms and functoriality lifting. There are a variety of partial results known; see for example [SV03, AGR93]. • We know the usual modular symbols span the top-degree cohomology for any arithmetic group Γ. Fix a class of generalized modular symbols by fixing the pair G, H and fixing some class of maps f . How much of the cohomology can one span for a general arithmetic group Γ ⊂ G(Q)? A simple example is given by the Ash–Borel construction for G = SL3 and H a Levi factor of a rational parabolic subgroup P of type (2, 1). In this case H ≃ SL2 (R) × R>0 and sits inside G via µ −1 ¶ α M 0 g g −1 , M ∈ SL2 (R), α ∈ R>0 , g ∈ SL3 (Q). 0 α For Γ ⊂ SL3 (Z) these symbols define a subspace S(2,1) ⊂ H 2 (Γ\X; C).

Are there Γ for which S(2,1) equals the full cohomology space? For general Γ how much is captured? Is there a nice combinatorial way to write down the relations among these classes? Can one cook up a generalization of Theorem A.22 for these classes and use it to compute Hecke eigenvalues?

Bibliography

[AB90]

A. Ash and A. Borel, Generalized modular symbols, Cohomology of arithmetic groups and automorphic forms (Luminy-Marseille, 1989), Springer, Berlin, 1990, pp. 57–75.

[ADT04]

Nadia Ben Atti and Gema M. D´ıaz-Toca, http://hlombardi.free.fr/ publis/ABMAvar.html (2004).

[AG00]

Avner Ash and Robert Gross, Generalized non-abelian reciprocity laws: a context for Wiles’ proof, Bull. London Math. Soc. 32 (2000), no. 4, 385– 397. MR 1760802 (2001h:11142)

[Aga00]

A. Agashe, The Birch and Swinnerton-Dyer formula for modular abelian varieties of analytic rank 0, Ph.D. thesis, University of California, Berkeley (2000).

[AGG84]

Avner Ash, Daniel Grayson, and Philip Green, Computations of cuspidal cohomology of congruence subgroups of SL(3, Z), J. Number Theory 19 (1984), no. 3, 412–436. MR 769792 (86g:11032)

[AGM]

Avner Ash, Paul E. Gunnells, and Mark McConnell, Cohomology of congruence subgroups of SL4 (Z) II, in preparation.

[AGM02]

, Cohomology of congruence subgroups of SL4 (Z), J. Number Theory 94 (2002), no. 1, 181–212. MR 1904968 (2003f:11072)

[AGR93]

Avner Ash, David Ginzburg, and Steven Rallis, Vanishing periods of cusp forms over modular symbols, Math. Ann. 296 (1993), no. 4, 709–723. MR 1233493 (94f:11044)

[Ahl78]

Lars V. Ahlfors, Complex analysis, third ed., McGraw-Hill Book Co., New York, 1978, An introduction to the theory of analytic functions of one complex variable, International Series in Pure and Applied Mathematics. MR 510197 (80c:30001)

[AL70]

A. O. L. Atkin and J. Lehner, Hecke operators on Γ0 (m), Math. Ann. 185 (1970), 134–160.

[AO01]

Scott Ahlgren and Ken Ono, Addition and counting: the arithmetic of partitions, Notices Amer. Math. Soc. 48 (2001), no. 9, 978–984. MR 1854533 (2002e:11136)

253

254

Bibliography

[AR79]

Avner Ash and Lee Rudolph, The modular symbol and continued fractions in higher dimensions, Invent. Math. 55 (1979), no. 3, 241–250. MR 553998 (82g:12011)

[Art79]

James Arthur, Eisenstein series and the trace formula, Automorphic forms, representations and L-functions (Proc. Sympos. Pure Math., Oregon State Univ., Corvallis, Ore., 1977), Part 1, Proc. Sympos. Pure Math., XXXIII, Amer. Math. Soc., Providence, R.I., 1979, pp. 253–274. MR 546601 (81b:10020)

[Ash77]

Avner Ash, Deformation retracts with lowest possible dimension of arithmetic quotients of self-adjoint homogeneous cones, Math. Ann. 225 (1977), no. 1, 69–76. MR 0427490 (55 #522)

[Ash80]

, Cohomology of congruence subgroups SL(n, Z), Math. Ann. 249 (1980), no. 1, 55–73. MR 82f:22010

[Ash84]

, Small-dimensional classifying spaces for arithmetic subgroups of general linear groups, Duke Math. J. 51 (1984), no. 2, 459–468. MR 747876 (85k:22027)

[Ash86]

, A note on minimal modular symbols, Proc. Amer. Math. Soc. 96 (1986), no. 3, 394–396. MR 822426 (87e:22024)

[Ash94]

, Unstable cohomology of SL(n, O), J. Algebra 167 (1994), no. 2, 330–342. MR 1283290 (95g:20050)

[Bar57]

E. S. Barnes, The perfect and extreme senary forms, Canad. J. Math. 9 (1957), 235–242. MR 0086834 (19,251e)

[Bar94]

A. Barvinok, A polynomial time algorithm for counting integral points in polyhedra when the dimension is fixed, Math. Oper. Res. 19 (1994), no. 4, 769–779.

[Bas96]

Jacques Basmaji, Ein Algorithmus zur Berechnung von Hecke-Operatoren und Anwendungen auf modulare Kurven, http://modular.math.washington.edu/scans/papers/basmaji/, 1996.

[BC06]

S. S. Bullock and C. Connell, Equivariant retracts of geometrically finite discrete groups acting on negatively pinched Hadamard manifolds, in preparation, 2006.

[BCDT01]

C. Breuil, B. Conrad, Fred Diamond, and R. Taylor, On the modularity of elliptic curves over Q: wild 3-adic exercises, J. Amer. Math. Soc. 14 (2001), no. 4, 843–939 (electronic). MR 2002d:11058

[BCP97]

W. Bosma, J. Cannon, and C. Playoust, The Magma algebra system. I. The user language, J. Symbolic Comput. 24 (1997), no. 3–4, 235–265, Computational algebra and number theory (London, 1993). MR 1 484 478

[BCS92]

J. P. Buhler, R. E. Crandall, and R. W. Sompolski, Irregular primes to one million, Math. Comp. 59 (1992), no. 200, 717–722. MR 1134717 (93a:11106)

[BHKS06]

K. Belebas, M. Van Hoeij, J. Kl¨ uners, and A. Steel, Factoring polynomials over global fields, preprint at http://www.math.fsu.edu/~hoeij/papers.html (2006).

[BI97]

R. Baeza and M. I. Icaza, On Humbert-Minkowski’s constant for a number field, Proc. Amer. Math. Soc. 125 (1997), no. 11, 3195–3202. MR 1403112 (97m:11092)

[Bir71]

B. J. Birch, Elliptic curves over Q: A progress report, 1969 Number Theory Institute (Proc. Sympos. Pure Math., Vol. XX, State Univ. New York,

Bibliography

255

Stony Brook, N.Y., 1969), Amer. Math. Soc., Providence, R.I., 1971, pp. 396–400. [BK90]

S. Bloch and K. Kato, L-functions and Tamagawa numbers of motives, The Grothendieck Festschrift, Vol. I, Birkh¨ auser Boston, Boston, MA, 1990, pp. 333–400.

[BMS06]

Yann Bugeaud, Maurice Mignotte, and Samir Siksek, Classical and modular approaches to exponential Diophantine equations. II. The LebesgueNagell equation, Compos. Math. 142 (2006), no. 1, 31–62. MR 2196761

[Bro94]

Kenneth S. Brown, Cohomology of groups, Graduate Texts in Mathematics, vol. 87, Springer-Verlag, New York, 1994, corrected reprint of the 1982 original. MR 1324339 (96a:20072)

[BS73]

A. Borel and J.-P. Serre, Corners and arithmetic groups, Comment. Math. Helv. 48 (1973), 436–491, avec un appendice: Arrondissement des vari´et´es ` a coins, par A. Douady et L. H´erault. MR 0387495 (52 #8337)

[BS02]

K. Buzzard and W. A. Stein, A mod five approach to modularity of icosahedral Galois representations, Pacific J. Math. 203 (2002), no. 2, 265–282. MR 2003c:11052

[BT82]

Raoul Bott and Loring W. Tu, Differential forms in algebraic topology, Graduate Texts in Mathematics, vol. 82, Springer-Verlag, New York, 1982. MR 658304 (83i:57016)

[Bul00]

S. S. Bullock, Well-rounded retracts of rank one symmetric spaces, preprint, 2000.

[Bum84]

Daniel Bump, Automorphic forms on GL(3, R), Lecture Notes in Mathematics, vol. 1083, Springer-Verlag, Berlin, 1984. MR 765698 (86g:11028)

[Bum97]

, Automorphic forms and representations, Cambridge Studies in Advanced Mathematics, vol. 55, Cambridge University Press, Cambridge, 1997. MR 1431508 (97k:11080)

[Buz96]

Kevin Buzzard, On the eigenvalues of the Hecke operator T2 , J. Number Theory 57 (1996), no. 1, 130–132. MR 96m:11033

[BW00]

A. Borel and N. Wallach, Continuous cohomology, discrete subgroups, and representations of reductive groups, second ed., Mathematical Surveys and Monographs, vol. 67, American Mathematical Society, Providence, RI, 2000. MR 1721403 (2000j:22015)

[Byg99]

J. Bygott, Modular forms and modular symbols over imaginary quadratic fields, Ph.D. thesis, Exeter University, 1999.

[Car59a]

L. Carlitz, Arithmetic properties of generalized Bernoulli numbers, J. Reine Angew. Math. 202 (1959), 174–182. MR 0109132 (22 #20)

[Car59b]

, Some arithmetic properties of generalized Bernoulli numbers, Bull. Amer. Math. Soc. 65 (1959), 68–69. MR 0104630 (21 #3383)

[CDT99]

Brian Conrad, Fred Diamond, and Richard Taylor, Modularity of certain potentially Barsotti-Tate Galois representations, J. Amer. Math. Soc. 12 (1999), no. 2, 521–567. MR 1639612 (99i:11037)

[CF67]

George E. Cooke and Ross L. Finney, Homology of cell complexes, Based on lectures by Norman E. Steenrod, Princeton University Press, Princeton, N.J., 1967. MR 0219059 (36 #2142)

[Che05]

Imin Chen, A Diophantine equation associated to X0 (5), LMS J. Comput. Math. 8 (2005), 116–121 (electronic). MR 2153792 (2006b:11052)

256

Bibliography

[CL04]

J. Cremona and M. P. Lingham, Finding all elliptic curves with good reduction outside a given set of primes, in progress (2004).

[CO77]

H. Cohen and J. Oesterl´e, Dimensions des espaces de formes modulaires, 69–78. Lecture Notes in Math., Vol. 627. MR 57 #12396

[Coh93]

H. Cohen, A course in computational algebraic number theory, SpringerVerlag, Berlin, 1993. MR 94i:11105

[Cou01]

Renaud Coulangeon, Vorono¨ı theory over algebraic number fields, R´eseaux euclidiens, designs sph´eriques et formes modulaires, Monogr. Enseign. Math., vol. 37, Enseignement Math., Geneva, 2001, pp. 147–162. MR 1878749 (2002m:11064)

[Cre]

J. E. Cremona, personal communication.

[Cre84]

, Hyperbolic tessellations, modular symbols, and elliptic curves over complex quadratic fields, Compositio Math. 51 (1984), no. 3, 275–324.

[Cre92]

, Modular symbols for Γ1 (N ) and elliptic curves with everywhere good reduction, Math. Proc. Cambridge Philos. Soc. 111 (1992), no. 2, 199–218.

[Cre97a]

, Algorithms for modular elliptic curves, second ed., Cambridge University Press, Cambridge, 1997, http://www.maths.nott.ac.uk/personal/jec/book/.

[Cre97b]

, Computing periods of cusp forms and modular elliptic curves, Experiment. Math. 6 (1997), no. 2, 97–107.

[Cre06]

, Proceedings of the 7th International Symposium (ANTS-VII) (2006).

[CS88]

J. H. Conway and N. J. A. Sloane, Low-dimensional lattices. III. Perfect forms, Proc. Roy. Soc. London Ser. A 418 (1988), no. 1854, 43–80. MR 953277 (90a:11073)

[CW94]

J. E. Cremona and E. Whitley, Periods of cusp forms and elliptic curves over imaginary quadratic fields, Math. Comp. 62 (1994), no. 205, 407–429.

[CWZ01]

Janos A. Csirik, Joseph L. Wetherell, and Michael E. Zieve, On the genera of X0 (N ), http://www.csirik.net/papers.html (2001).

[Dar97]

H. Darmon, Faltings plus epsilon, Wiles plus epsilon, and the generalized Fermat equation, C. R. Math. Rep. Acad. Sci. Canada 19 (1997), no. 1, 3–14. MR 1479291 (98h:11034a)

[Dem04]

L. Demb´el´e, Quaternionic Manin symbols, Brandt matrices and Hilbert modular forms, preprint, 2004. √ L. Demb´el´e, Explicit computations of Hilbert modular forms on Q( 5), Experiment. Math. 14 (2005), no. 4, 457–466. MR 2193808

[Dem05] [DI95]

F. Diamond and J. Im, Modular forms and modular curves, Seminar on Fermat’s Last Theorem, Providence, RI, 1995, pp. 39–133.

[Dia96]

F. Diamond, On deformation rings and Hecke rings, Ann. of Math. (2) 144 (1996), no. 1, 137–166. MR 1405946 (97d:11172)

[Dix82]

John D. Dixon, Exact solution of linear equations using p-adic expansions, Numer. Math. 40 (1982), no. 1, 137–141. MR 681819 (83m:65025)

[Dok04]

Tim Dokchitser, Computing special values of motivic L-functions, Experiment. Math. 13 (2004), no. 2, 137–149.

[DP04]

H. Darmon and R. Pollack, The efficient calculation of Stark-Heegner points via overconvergent modular symbols.

Bibliography

257

[DS05]

Fred Diamond and Jerry Shurman, A first course in modular forms, Graduate Texts in Mathematics, vol. 228, Springer-Verlag, New York, 2005.

[DVS05]

M. Dutour, F. Vallentin, and A. Sch¨ urmann, Classification of perfect forms in dimension 8, talk at Oberwolfach meeting Sphere packings: Exceptional structures and relations to other fields, November 2005.

[Ebe02]

Wolfgang Ebeling, Lattices and codes, revised ed., Advanced Lectures in Mathematics, Friedr. Vieweg & Sohn, Braunschweig, 2002, a course partially based on lectures by F. Hirzebruch.

[ECdJ+ 06]

Bas Edixhoven, Jean-Marc Couveignes, Robin de Jong, Franz Merkl, and Johan Bosman, On the computation of coefficients of modular form, http://www.arxiv.org/abs/math.NT/0605244 (2006).

[EGM98]

J. Elstrodt, F. Grunewald, and J. Mennicke, Groups acting on hyperbolic space, Springer Monographs in Mathematics, Springer-Verlag, Berlin, 1998, Harmonic analysis and number theory. MR 1483315 (98g:11058)

[Eil47]

Samuel Eilenberg, Homology of spaces with operators. I, Trans. Amer. Math. Soc. 61 (1947), 378–417; errata, 62, 548 (1947). MR 0021313 (9,52b)

[Elk98]

Noam D. Elkies, Elliptic and modular curves over finite fields and related computational issues, Computational perspectives on number theory (Chicago, IL, 1995), AMS/IP Stud. Adv. Math., vol. 7, Amer. Math. Soc., Providence, RI, 1998, pp. 21–76. MR 1486831 (99a:11078)

[EVGS02]

Philippe Elbaz-Vincent, Herbert Gangl, and Christophe Soul´e, Quelques calculs de la cohomologie de GLN (Z) et de la K-th´eorie de Z, C. R. Math. Acad. Sci. Paris 335 (2002), no. 4, 321–324. MR 1931508 (2003h:19002)

[FH91]

William Fulton and Joe Harris, Representation theory, Graduate Texts in Mathematics, vol. 129, Springer-Verlag, New York, 1991, A first course, Readings in Mathematics. MR 1153249 (93a:20069)

[FJ02]

D. W. Farmer and K. James, The irreducibility of some level 1 Hecke polynomials, Math. Comp. 71 (2002), no. 239, 1263–1270 (electronic). MR 2003e:11046

[FL]

D. W. Farmer and Stefan Lemurell, Maass forms and their L-functions, AIM 2005-15, arXiv:math.NT/0506102.

[FM99]

G. Frey and M. M¨ uller, Arithmetic of modular curves and applications, Algorithmic algebra and number theory (Heidelberg, 1997), Springer, Berlin, 1999, pp. 11–48. ´ J. Franke, Harmonic analysis in weighted L2 -spaces, Ann. Sci. Ecole Norm. Sup. (4) 31 (1998), no. 2, 181–279.

[Fra98] [FT93]

A. Fr¨ ohlich and M. J. Taylor, Algebraic number theory, Cambridge University Press, Cambridge, 1993.

[FvdG]

C. Faber and G. van der Geer, Sur la cohomologie des Syst´emes Locaux sur les Espaces des Modules des Courbes de Genus 2 and des Surfaces Ab´eliennes, arXiv:math.AG/0305094.

[Gel75]

Stephen S. Gelbart, Automorphic forms on ad`ele groups, Princeton University Press, Princeton, N.J., 1975, Annals of Mathematics Studies, No. 83. MR 0379375 (52 #280)

[GH81]

M. J. Greenberg and J. R. Harper, Algebraic topology, Benjamin/Cummings Publishing Co. Inc. Advanced Book Program, Reading, Mass., 1981, A first course. MR 83b:55001

258

Bibliography

[GLQ04]

Josep Gonz´ alez, Joan-Carles Lario, and Jordi Quer, Arithmetic of Qcurves, Modular curves and abelian varieties, Progr. Math., vol. 224, Birkh¨ auser, Basel, 2004, pp. 125–139. MR 2058647 (2005c:11068)

[GM03]

P. E. Gunnells and M. McConnell, Hecke operators and Q-groups associated to self-adjoint homogeneous cones, J. Number Theory 100 (2003), no. 1, 46–71.

[Gol05]

Dorian Goldfeld, Automorphic forms and L-functions on the general linear group, to appear, 2005.

[Gon97]

A. B. Goncharov, The double logarithm and Manin’s complex for modular curves, Math. Res. Lett. 4 (1997), no. 5, 617–636.

[Gon98]

, Multiple polylogarithms, cyclotomy and modular complexes, Math. Res. Lett. 5 (1998), no. 4, 497–516.

[Gor93]

D. Gordon, Discrete logarithms in GF(p) using the number field sieve, SIAM J. Discrete Math. 6 (1993), no. 1, 124–138. MR 94d:11104

[Gor04]

, Discrete logarithm problem, http://www.win.tue.nl/~henkvt/content.html.

[GP05]

Benedict H. Gross and David Pollack, On the Euler characteristic of the discrete spectrum, J. Number Theory 110 (2005), no. 1, 136–163. MR 2114678 (2005k:11100)

[Gre83]

Ralph Greenberg, On the Birch and Swinnerton-Dyer conjecture, Invent. Math. 72 (1983), no. 2, 241–265. MR 700770 (85c:11052)

[Gri05]

G. Grigorov, Kato’s Euler System and the Main Conjecture, Harvard Ph.D. Thesis (2005).

[Gro98]

Benedict H. Gross, On the Satake isomorphism, Galois representations in arithmetic algebraic geometry (Durham, 1996), London Math. Soc. Lecture Note Ser., vol. 254, Cambridge Univ. Press, Cambridge, 1998, pp. 223–237. MR 1696481 (2000e:22008)

[GS81]

F. Grunewald and J. Schwermer, A nonvanishing theorem for the cuspidal cohomology of SL2 over imaginary quadratic integers, Math. Ann. 258 (1981), 183–200.

[GS02]

Mark Giesbrecht and Arne Storjohann, Computing rational forms of integer matrices, J. Symbolic Comput. 34 (2002), no. 3, 157–172. MR 1935075 (2003j:15016)

[Gun99]

P. E. Gunnells, Modular symbols for Q-rank one groups and Vorono˘ı reduction, J. Number Theory 75 (1999), no. 2, 198–219.

[Gun00a]

, Computing Hecke eigenvalues below the cohomological dimension, Experiment. Math. 9 (2000), no. 3, 351–367. MR 1 795 307

[Gun00b]

, Symplectic modular symbols, Duke Math. J. 102 (2000), no. 2, 329–350.

[Hara]

G. Harder, Congruences between modular forms of genus 1 and of genus 2, Arbeitstagung.

[Harb]

, Kohomologie arithmetischer Gruppen, lecture notes, Universit¨ at Bonn, 1987–1988.

[Har87]

, Eisenstein cohomology of arithmetic groups. The case GL2 , Invent. Math. 89 (1987), no. 1, 37–118. MR 892187 (89b:22018)

Bibliography

259

[Har91]

, Eisenstein cohomology of arithmetic groups and its applications to number theory, Proceedings of the International Congress of Mathematicians, Vol. I, II (Kyoto, 1990) (Tokyo), Math. Soc. Japan, 1991, pp. 779– 790. MR 1159264 (93b:11057)

[Har05]

, Modular symbols and special values of automorphic L-functions, preprint, 2005.

[HC68]

Harish-Chandra, Automorphic forms on semisimple Lie groups, Notes by J. G. M. Mars. Lecture Notes in Mathematics, No. 62, Springer-Verlag, Berlin, 1968. MR 0232893 (38 #1216)

[Hel01]

Sigurdur Helgason, Differential geometry, Lie groups, and symmetric spaces, Graduate Studies in Mathematics, vol. 34, American Mathematical Society, Providence, RI, 2001, corrected reprint of the 1978 original. MR 1834454 (2002b:53081)

[Hij74]

H. Hijikata, Explicit formula of the traces of Hecke operators for Γ0 (N ), J. Math. Soc. Japan 26 (1974), no. 1, 56–82.

[Hsu96]

Tim Hsu, Identifying congruence subgroups of the modular group, Proc. Amer. Math. Soc. 124 (1996), no. 5, 1351–1359. MR 1343700 (96k:20100)

[HT01]

Michael Harris and Richard Taylor, The geometry and cohomology of some simple Shimura varieties, Annals of Mathematics Studies, vol. 151, Princeton University Press, Princeton, NJ, 2001, with an appendix by Vladimir G. Berkovich. MR 1876802 (2002m:11050)

[Hum80]

James E. Humphreys, Arithmetic groups, Lecture Notes in Mathematics, vol. 789, Springer, Berlin, 1980. MR 584623 (82j:10041)

[Ica97]

M. I. Icaza, Hermite constant and extreme forms for algebraic number fields, J. London Math. Soc. (2) 55 (1997), no. 1, 11–22. MR 1423282 (97j:11034)

[Jaq91]

David-Olivier Jaquet, Classification des r´eseaux dans R7 (via la notion de formes parfaites), Ast´erisque (1991), no. 198-200, 7–8, 177–185 (1992), Journ´ees Arithm´etiques, 1989 (Luminy, 1989). MR 1144322 (93g:11071)

[JBS03]

A. Jorza, J. Balakrishna, and W. Stein, The Smallest Conductor for an Elliptic Curve of Rank Four is Composite, http://modular.math.washington.edu/rank4/. ´ David-Olivier Jaquet-Chiffelle, Enum´ eration compl`ete des classes de formes parfaites en dimension 7, Ann. Inst. Fourier (Grenoble) 43 (1993), no. 1, 21–55. MR 1209694 (94d:11048)

[JC93]

[Kan00]

Masanobu Kaneko, The Akiyama-Tanigawa algorithm for Bernoulli numbers, J. Integer Seq. 3 (2000), no. 2, Article 00.2.9, 6 pp. (electronic). MR 1800883 (2001k:11026)

[Kel06]

Bernd C. Kellner, Bernoulli numbers, http://www.bernoulli.org (2006).

[Kna92]

A. W. Knapp, Elliptic curves, Princeton University Press, Princeton, NJ, 1992.

[Knu]

Donald E. Knuth, The art of computer programming. Vol. 2, third ed., Addison-Wesley Publishing Co., Reading, Mass., Seminumerical algorithms, Addison-Wesley Series in Computer Science and Information Processing.

[Kob84]

N. Koblitz, Introduction to elliptic curves and modular forms, Graduate Texts in Mathematics, vol. 97, Springer-Verlag, New York, 1984. MR 86c:11040

260

Bibliography

[Kri90]

Aloys Krieg, Hecke algebras, Mem. Amer. Math. Soc. 87 (1990), no. 435, x+158. MR 1027069 (90m:16024)

[Laf02]

Laurent Lafforgue, Chtoucas de Drinfeld et correspondance de Langlands, Invent. Math. 147 (2002), no. 1, 1–241. MR 1875184 (2002m:11039)

[Lan66]

R. P. Langlands, Eisenstein series, Algebraic Groups and Discontinuous Subgroups (Proc. Sympos. Pure Math., Boulder, Colo., 1965), Amer. Math. Soc., Providence, R.I., 1966, pp. 235–252. MR 0249539 (40 #2784)

[Lan76]

Robert P. Langlands, On the functional equations satisfied by Eisenstein series, Springer-Verlag, Berlin, 1976, Lecture Notes in Mathematics, Vol. 544. MR 0579181 (58 #28319)

[Lan95]

S. Lang, Introduction to modular forms, Springer-Verlag, Berlin, 1995, with appendixes by D. Zagier and W. Feit, corrected reprint of the 1976 original.

[Lem01]

Dominic Lemelin, Mazur-tate type conjectures for elliptic curves defined over quadratic imaginary fields.

[Leo58]

Heinrich-Wolfgang Leopoldt, Eine Verallgemeinerung der Bernoullischen Zahlen, Abh. Math. Sem. Univ. Hamburg 22 (1958), 131–140. MR 0092812 (19,1161e)

[Li75]

W-C. Li, Newforms and functional equations, Math. Ann. 212 (1975), 285–315.

[Lin05]

M. Lingham, Modular forms and elliptic curves over imaginary quadratic fields, Ph.D. thesis, Nottingham, 2005.

[LLL82]

A. K. Lenstra, H. W. Lenstra, Jr., and L. Lov´ asz, Factoring polynomials with rational coefficients, Math. Ann. 261 (1982), no. 4, 515–534. MR 682664 (84a:12002)

[LS76]

Ronnie Lee and R. H. Szczarba, On the homology and cohomology of congruence subgroups, Invent. Math. 33 (1976), no. 1, 15–53. MR 0422498 (54 #10485)

[LS90]

J.-P. Labesse and J. Schwermer (eds.), Cohomology of arithmetic groups and automorphic forms, Lecture Notes in Mathematics, vol. 1447, Berlin, Springer-Verlag, 1990. MR 1082959 (91h:11033)

[LS02]

Joan-C. Lario and Ren´e Schoof, Some computations with Hecke rings and deformation rings, Experiment. Math. 11 (2002), no. 2, 303–311, with an appendix by Amod Agashe and William Stein. MR 1959271 (2004b:11072)

[LS04]

Jian-Shu Li and Joachim Schwermer, On the Eisenstein cohomology of arithmetic groups, Duke Math. J. 123 (2004), no. 1, 141–169. MR 2060025 (2005h:11108)

[Lub94]

A. Lubotzky, Discrete groups, expanding graphs and invariant measures, Progress in Mathematics, vol. 125, Birkh¨ auser Verlag, Basel, 1994, with an appendix by Jonathan D. Rogawski.

[Man72]

J. I. Manin, Parabolic points and zeta functions of modular curves, Izv. Akad. Nauk SSSR Ser. Mat. 36 (1972), 19–66. MR 47 #3396

[Mar01]

Fran¸cois Martin, P´eriodes de formes modulaires de poids 1.

[Mar03]

Jacques Martinet, Perfect lattices in Euclidean spaces, Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences], vol. 327, Springer-Verlag, Berlin, 2003. MR 1957723 (2003m:11099)

Bibliography

261

[Mar05]

Greg Martin, Dimensions of the spaces of cusp forms and newforms on Γ0 (N ) and Γ1 (N ), J. Number Theory 112 (2005), no. 2, 298–331. MR 2141534 (2005m:11069)

[Maz73]

B. Mazur, Courbes elliptiques et symboles modulaires, S´eminaire Bourbaki, 24`eme ann´ee (1971/1972), Exp. No. 414, Springer, Berlin, 1973, pp. 277– 294. Lecture Notes in Math., Vol. 317. MR 55 #2930

[McC91]

M. McConnell, Classical projective geometry and arithmetic groups, Math. Ann. 290 (1991), no. 3, 441–462. MR 92k:22020

[Men79]

Eduardo R. Mendoza, Cohomology of PGL2 over imaginary quadratic integers, Bonner Mathematische Schriften [Bonn Mathematical Publications], 128, Universit¨ at Bonn Mathematisches Institut, Bonn, 1979, Dissertation, Rheinische Friedrich-Wilhelms-Universit¨ at, Bonn, 1979. MR 611515 (82g:22012)

[Mer94]

L. Merel, Universal Fourier expansions of modular forms, On Artin’s conjecture for odd 2-dimensional representations, Springer, 1994, pp. 59–94.

[Mer99]

, Arithmetic of elliptic curves and Diophantine equations, J. Th´eor. Nombres Bordeaux 11 (1999), no. 1, 173–200, Les XX`emes Journ´ees Arithm´etiques (Limoges, 1997). MR 1730439 (2000j:11084)

[Mes86]

J.-F. Mestre, La m´ethode des graphes. Exemples et applications, Proceedings of the international conference on class numbers and fundamental units of algebraic number fields (Katata) (1986), 217–242.

[Miy89]

T. Miyake, Modular forms, Springer-Verlag, Berlin, 1989, translated from the Japanese by Yoshitaka Maeda.

[MM89]

R. MacPherson and M. McConnell, Classical projective geometry and modular varieties, Algebraic analysis, geometry, and number theory (Baltimore, MD, 1988), Johns Hopkins Univ. Press, Baltimore, MD, 1989, pp. 237–290. MR 98k:14076

[MM93]

, Explicit reduction theory for Siegel modular threefolds, Invent. Math. 111 (1993), no. 3, 575–625. MR 94a:32052

[MTT86]

B. Mazur, J. Tate, and J. Teitelbaum, On p-adic analogues of the conjectures of Birch and Swinnerton-Dyer, Invent. Math. 84 (1986), no. 1, 1–48.

[MW94]

Colette Mœglin and Jean-Loup Waldspurger, D´ecomposition spectrale et s´eries d’Eisenstein, Progress in Mathematics, vol. 113, Birkh¨ auser Verlag, ´ Basel, 1994, Une paraphrase de l’Ecriture [A paraphrase of Scripture]. MR 1261867 (95d:11067)

[Nec94]

V. I. Nechaev, On the complexity of a deterministic algorithm for a discrete logarithm, Mat. Zametki 55 (1994), no. 2, 91–101, 189. MR 96a:11145

[Ong86]

Heidrun E. Ong, Perfect quadratic forms over real-quadratic number fields, Geom. Dedicata 20 (1986), no. 1, 51–77. MR 823160 (87f:11023)

[PR94]

Vladimir Platonov and Andrei Rapinchuk, Algebraic groups and number theory, Pure and Applied Mathematics, vol. 139, Academic Press Inc., Boston, MA, 1994, translated from the 1991 Russian original by Rachel Rowen. MR 1278263 (95b:11039)

[Que06]

J. Quer, Dimensions of spaces of modular forms for ΓH (N ), Preprint.

[Rib92]

K. A. Ribet, Abelian varieties over Q and modular forms, Algebra and topology 1992 (Taej˘ on), Korea Adv. Inst. Sci. Tech., Taej˘ on, 1992, pp. 53– 79. MR 94g:11042

262

[Ros86] [RS01]

[Sap97]

[Sar03] [SC03] [Sch86]

[Sch90] [Sch95]

[Ser73] [Ser87] [Shi59] [Shi94]

[Sho80a] [Sho80b] [Sho97]

[Sil92] [Sou75] [Sta79] [Ste] [Ste97]

Bibliography

M. Rosen, Abelian varieties over C, Arithmetic geometry (Storrs, Conn., 1984), Springer, New York, 1986, pp. 79–101. K. A. Ribet and W. A. Stein, Lectures on Serre’s conjectures, Arithmetic algebraic geometry (Park City, UT, 1999), IAS/Park City Math. Ser., vol. 9, Amer. Math. Soc., Providence, RI, 2001, pp. 143–232. MR 2002h:11047 Leslie Saper, Tilings and finite energy retractions of locally symmetric spaces, Comment. Math. Helv. 72 (1997), no. 2, 167–202. MR 1470087 (99a:22019) Peter Sarnak, Spectra of hyperbolic surfaces, Bull. Amer. Math. Soc. (N.S.) 40 (2003), no. 4, 441–478 (electronic). MR 1997348 (2004f:11107) Samir Siksek and John E. Cremona, On the Diophantine equation x2 + 7 = y m , Acta Arith. 109 (2003), no. 2, 143–149. MR 1980642 (2004c:11109) Joachim Schwermer, Holomorphy of Eisenstein series at special points and cohomology of arithmetic subgroups of SLn (Q), J. Reine Angew. Math. 364 (1986), 193–220. MR 817646 (87h:11048) A. J. Scholl, Motives for modular forms, Invent. Math. 100 (1990), no. 2, 419–430. R. Schoof, Counting points on elliptic curves over finite fields, J. Th´eor. Nombres Bordeaux 7 (1995), no. 1, 219–254, Les Dix-huiti`emes Journ´ees Arithm´etiques (Bordeaux, 1993). MR 1413578 (97i:11070) J-P. Serre, A Course in Arithmetic, Springer-Verlag, New York, 1973, Translated from the French, Graduate Texts in Mathematics, No. 7. , Sur les repr´esentations modulaires de degr´e 2 de Gal(Q/Q), Duke Math. J. 54 (1987), no. 1, 179–230. G. Shimura, Sur les int´egrales attach´ees aux formes automorphes, J. Math. Soc. Japan 11 (1959), 291–311. , Introduction to the arithmetic theory of automorphic functions, Princeton University Press, Princeton, NJ, 1994, reprint of the 1971 original, Kan Memorial Lectures, 1. V. V. Shokurov, Shimura integrals of cusp forms, Izv. Akad. Nauk SSSR Ser. Mat. 44 (1980), no. 3, 670–718, 720. MR 582162 (82b:10029) , A study of the homology of Kuga varieties, Izv. Akad. Nauk SSSR Ser. Mat. 44 (1980), no. 2, 443–464, 480. MR 571104 (82f:14023) Victor Shoup, Lower bounds for discrete logarithms and related problems, Advances in cryptology—EUROCRYPT ’97 (Konstanz), Lecture Notes in Comput. Sci., vol. 1233, Springer, Berlin, 1997, pp. 256–266. MR 98j:94023 J. H. Silverman, The arithmetic of elliptic curves, Springer-Verlag, New York, 1992, corrected reprint of the 1986 original. Christophe Soul´e, Cohomologie de SL3 (Z), C. R. Acad. Sci. Paris S´er. A-B 280 (1975), no. 5, Ai, A251–A254. MR 0396849 (53 #709) R. E. Staffeldt, Reduction theory and K3 of the Gaussian integers, Duke Math. J. 46 (1979), no. 4, 773–798. MR 552526 (80m:22014) Allan Steel, Advanced matrix algorithms, Seminar Talk at Harvard University. , A new algorithm for the computation of canonical forms of matrices over fields, J. Symbolic Comput. 24 (1997), no. 3-4, 409–432, Computational algebra and number theory (London, 1993). MR 1484489 (98m:65070)

263

Bibliography

[Ste99a]

Norman Steenrod, The topology of fibre bundles, Princeton Landmarks in Mathematics, Princeton University Press, Princeton, NJ, 1999, reprint of the 1957 edition, Princeton Paperbacks. MR 1688579 (2000a:55001)

[Ste99b]

W. A. Stein, HECKE: The Modular Symbols Calculator, software (available online) (1999).

[Ste00]

, Explicit approaches to modular abelian varieties, Ph.D. thesis, University of California, Berkeley (2000).

[Ste06]

, SAGE: Software for Algebra and Geometry Experimentation, http://sage.scipy.org/sage.

[Str69]

Volker Strassen, Gaussian elimination is not optimal, Numerische Mathematik 13 (1969), 354–356.

[Stu87]

J. Sturm, On the congruence of modular forms, Number theory (New York, 1984–1985), Springer, Berlin, 1987, pp. 275–280.

[SV01]

W. A. Stein and H. A. Verrill, Cuspidal modular symbols are transportable, LMS J. Comput. Math. 4 (2001), 170–181 (electronic). MR 1 901 355

[SV03]

B. Speh and T. N. Venkataramana, Construction of some generalised modular symbols, preprint, 2003.

[SW02]

William A. Stein and Mark Watkins, A database of elliptic curves—first report, Algorithmic number theory (Sydney, 2002), Lecture Notes in Comput. Sci., vol. 2369, Springer, Berlin, 2002, pp. 267–275. MR 2041090 (2005h:11113)

[SW05]

Jude Socrates and David Whitehouse, Unramified Hilbert modular forms, with examples relating to elliptic curves, Pacific J. Math. 219 (2005), no. 2, 333–364. MR 2175121

[Tat75]

J. Tate, Algorithm for determining the type of a singular fiber in an elliptic pencil, Modular functions of one variable, IV (Proc. Internat. Summer School, Univ. Antwerp, Antwerp, 1972), Springer, Berlin, 1975, pp. 33–52. Lecture Notes in Math., Vol. 476. MR 52 #13850

[Tho89]

J. G. Thompson, Hecke operators and noncongruence subgroups, Group theory (Singapore, 1987), de Gruyter, Berlin, 1989, including a letter from J.-P. Serre, pp. 215–224. MR 981844 (90a:20105)

[Tot05]

A. Toth, On the Steinberg module of Chevalley groups, Manuscripta Math. 116 (2005), no. 3, 277–295.

[TW95]

R. Taylor and A. J. Wiles, Ring-theoretic properties of certain Hecke algebras, Ann. of Math. (2) 141 (1995), no. 3, 553–572.

[vdG]

Gerard van der Geer, Siegel Modular Forms, arXiv:math.AG/0605346.

[vGvdKTV97] Bert van Geemen, Wilberd van der Kallen, Jaap Top, and Alain Verberkmoes, Hecke eigenforms in the cohomology of congruence subgroups of SL(3, Z), Experiment. Math. 6 (1997), no. 2, 163–174. MR 1474576 (99a:11059) [Vig77]

Marie-France Vign´eras, S´eries thˆeta des formes quadratiques ind´efinies, S´eminaire Delange-Pisot-Poitou, 17e ann´ee (1975/76), Th´eorie des nombres: Fasc. 1, Exp. No. 20, Secr´etariat Math., Paris, 1977, p. 3. MR 0480352 (58 #521)

[Vog85]

K. Vogtmann, Rational homology of Bianchi groups, Math. Ann. 272 (1985), no. 3, 399–419.

264

Bibliography

[Vog97]

David A. Vogan, Jr., Cohomology and group representations, Representation theory and automorphic forms (Edinburgh, 1996), Proc. Sympos. Pure Math., vol. 61, Amer. Math. Soc., Providence, RI, 1997, pp. 219–243. MR 1476500 (98k:22064)

[Vor08]

G. Voronoˇı, Nouvelles applications des param´etres continus ` a la th´eorie des formes quadratiques, I. Sur quelques propri´et´es des formes quadratiques positives parfaites, J. Reine Angew. Math. 133 (1908), 97–178.

[VZ84]

David A. Vogan, Jr. and Gregg J. Zuckerman, Unitary representations with nonzero cohomology, Compositio Math. 53 (1984), no. 1, 51–90. MR 762307 (86k:22040)

[Wan82]

Kai Wang, A proof of an identity of the Dirichlet L-function, Bull. Inst. Math. Acad. Sinica 10 (1982), no. 3, 317–321. MR 679019 (84c:10040)

[Wan95]

Xiang Dong Wang, 2-dimensional simple factors of J0 (N ), Manuscripta Math. 87 (1995), no. 2, 179–197. MR 1334940 (96h:11059)

[Wes]

U. Weselman, personal communication.

[Whi90]

E. Whitley, Modular symbols and elliptic curves over imaginary quadratic number fields, Ph.D. thesis, Exeter University, 1990.

[Wie05]

Gabor Wiese, Modular Forms of Weight One Over Finite Fields, Ph.D. thesis (2005).

[Wil95]

A. J. Wiles, Modular elliptic curves and Fermat’s last theorem, Ann. of Math. (2) 141 (1995), no. 3, 443–551. MR 1333035 (96d:11071)

[Wil00]

, The Birch and Swinnerton-Dyer Conjecture, http://www.claymath.org/prize problems/birchsd.htm.

[Yas05a]

D. Yasaki, On the cohomology of SU(2, 1) over the Gaussian integers, preprint, 2005.

[Yas05b]

, On the existence of spines for Q-rank 1 groups, preprint, 2005.

Index

Symbol Index C(Γ), 5 C[[q]], 4 ∆, 15 ε(γ), 180 F , 17 f [γ]k , 5 Γ(N ), 4 Γ0 (N ), 5 Γ1 (N ), 4 Gk (z), 13 GL2 (Q), 5 h, 1 h∗ , 6 j-function, 170 Mat2 (Z)n , 131 Mk (G), 123 Mk (G; R), 124 Mk (Γ), 7 Mk (N, ε), Mk (N, ε), 128 Mk , 17 Mk , 179 P1 (Q), 5 Sk (Γ), 134 Sk , 18 SL2 (Z), 1, 5 180

Algorithm Index p-adic Nullspace, 118 145 146 187 65 Asymptotically Fast Echelon Form, 111 Baby-step Giant-step Discrete Log, 69 Basis for Mk , 19 Basis of Cusp Forms, 56 Berlekamp-Massey, 116

Bernoulli Number Bn , 32 Conductor, 71 Cremona’s Heilbronn Matrices, 48 Cusp Representation, 135 Decomposition Using Kernels, 119 Dirichlet Character as Kronecker Symbol, 74 Elliptic Curves of Conductor N , Enumerating Eisenstein Series, 88 Evaluate ε, 68 Explicit Cusp Equivalence, 135 Extension of Character, 76 Factorization of Character, 71 Galois Orbit, 76 Gauss Elimination, 104 Generalized Bernoulli Numbers, 84 Hecke Operator, 26 Kronecker Symbol as Dirichlet Character, 74 List P1 (Z/N Z), Merel’s Algorithm for Computing a Basis, 165 Minimal Generator for (Z/pr Z)∗ , Modular Symbols Presentation, 154 Multimodular Echelon Form, 107 Order of Character, 70 Period Integrals, 181 Rational Reconstruction, 106 Reduction in P1 (Z/N Z) to Canonical Form, Restriction of Character, 75 Sum over A4 (N ), 99 System of Eigenvalues, 166 Values of ε, 70 Width of Cusp, 9

265

266

Definition Index Γ-invariant on the left, 206 k-sharblies, 233 q-expansion, 4 Q-rank, 245 123 125 128 13 134 159 212 218 226 7 abelian variety attached to f , 178 action of Hecke operators, 139 antiholomorphic, 137 arithmetic group, 208 associate proper Q-parabolic subgroups of G, automorphic form, 209 automorphy factor, 205 Bernoulli numbers, 16 Bianchi groups, 248 Borel conjecture, 212 boundary map, 40, 134 bounded domains, 211 cellular decomposition, 219 character of the modular form, 160 Cholesky decomposition, 214 codimension, 219 complex upper half plane, 1 conductor, 71 congruence subgroup, 4, 208 congruence subgroup problem, 7 Connected, 207 critical integers, 138 cross polytope, 238 cusp form, 4 cuspidal, 209 cuspidal automorphic form, 210 cuspidal cohomology, 212 cuspidal modular symbols, 40, cusps for a congruence subgroup Γ, 5 Defined over Q, 207 degeneracy map, 59, 161 diamond-bracket action, 160 diamond-bracket operators, 128, dimension, 219 Dirichlet character, 64 divisor, xi echelon form, 103 eigenforms, 59 Eilenberg–Mac Lane, 211 Eisenstein cohomology, 212 Eisenstein series, 210 Eisenstein subspace, 83 extended modular symbols, 179 extended upper half plane, 6 fan, 218 Farey tessellation, 220 formal power series, 4 Fourier expansion, 3 generalized Bernoulli numbers, 83 generalized modular symbol, 251

Index

Grothendieck motive, 179 group cohomology, 211 Hecke algebra, 54, 83, Hecke correspondence, 225 Hecke operator, 37, 128, Hecke polynomials, 241 height, 107 Hermite normal form, 120, 240 Hermitian symmetric spaces, 211 holomorphic, 2 holomorphic at ∞, 4 holomorphic at the cusp α, Humbert forms, 244 hypersimplices, 246 Krylov methods, 116 Krylov subspace, 116 Laplace–Beltrami–Casimir operator, 209 left action of G, 123 left action of GL2 (Q), 40 left action of SL2 (Z), 133 left translations, 208 level 1, 4 level of Γ, 4 linear fractional transformations, 1 Maass forms, 210 Manin symbol, 124 meromorphic, 2 meromorphic at ∞, 4 Miller basis, 20 modular complex, 244 modular elliptic curves, 187 modular form, 4, 7 modular function, 4 modular group, 2 modular symbols, 228 modular symbols algorithm, 229 modular symbols for Γ0 (N ), 40 modular symbols over a ring R, 124 newform, 59, 164 new modular symbols, 143 new subspace, 59, 162 nonnormalized weight k Eisenstein series, normalized Eisenstein series, 17 old modular symbols, 144 old subspace, 161 opposite, 222 perfect, 216 perfection, 244 pivot column, 103 plus one quotient, 165 primitive, 71, 215 primitive character associated to, 71 principal congruence subgroup, 208 Ramanujan function, 25 rational Jordan form, 114 rational period mapping, 185 real-analytic, 210

267

Index

reduced, 234 reducing point, 230 regular, 219 relative to the cusps, 39 restriction of scalars, 207 right action of SL2 (Z), 44, right translation, 209 satisfies condition Cn , 131 self-adjoint homogeneous cone, 248 Semisimple, 207 set of cusps, 5 Set of real points, 207 sharbly complex, 233 sigma function, 15 slowly increasing, 209 split form of SLn , 207 split symplectic group, 208 standard fundamental domain, 17 star involution, 141 strong deformation retract, 219 symplectic sharbly complex, 250 tilings, 246 topological cell, 218 transportable, 182 unimodular, 229 virtual cohomological dimension, 215 Voronoˇı decomposition, 219 Voronoˇı polyhedron, 215 Voronoˇı reduction algorithm, weakly modular function, 3, 5 Weierstrass ℘-function, 14 weight, 3, 4, 7 weight k modular symbols for G, weight k right action, 5 well-rounded retract, 219 width of the cusp, 6, 8

SAGE Index SAGE , vii, x, 2, M36 , 28 q-expansion of ∆, 15 SL2 (Z), 2 Z/N Z, 65 106, 144, 161, 15, 16, 20, 22, 163, 198 26, 30, 41, 43, 45, 51, 52, 49 50 56, 58, 63, 65–67, 74, 77, 78, 85, 89, 95, basis for M24 , 20 basis for S2 (Γ0 (N )), 56 Bernoulli numbers, 16 Bernoulli numbers modulo p, 30 boundary map, 52 continued fraction convergents, 43 cuspidal submodule, 52 dimension formulas, 93 dimension Sk (Γ0 (N )), 95 dimension Sk (Γ1 (N )), 97

dimension with character, 101, 161 Dirichlet character tutorial, 78 Dirichlet group, 67 echelon form, 112 Eisenstein arithmetic, 26 Eisenstein series, 89 evaluation of character, 67 generalized Bernoulli numbers, 85 Hecke operators M2 (Γ0 (39)), Hecke operators M2 (Γ0 (6)), Hecke operator T2 , 49 Heilbronn matrices, 49 Manin symbols, 45 Miller basis, 22 modular symbols, 44 modular symbols of level 11, 41 modular symbols printing, 46 rational reconstruction, 106

General Index 148, 150 SAGE, 49 Basmaji’s trick, 133 Bernoulli numbers generalized, 83 Birch and Swinnerton-Dyer conjecture, 10 boundary map, 134 computing, 51 boundary modular symbols and Manin symbols, 134 congruent number problem, 10 conjecture Maeda, 28 Shimura-Taniyama, 37 cusp forms ∆, 14 for Γ, 134 higher level dimension, 92, 96 cuspidal modular symbols and Manin symbols, 134 cusps action of SL2 (Z) on, 5 and boundary map, 134 criterion for vanishing, 136 dimension cusp forms of higher level, 92, 96 Diophantine equations, 10 Dirichlet character, 142 and cusps, 136 Eisenstein series, 13 algorithm to enumerate, 88 and Bernoulli numbers, 83 are eigenforms, 88 basis of, 88 compute, 63 compute using SAGE, 89

268

Fourier expansion, 15 Eisenstein subspace, 83 Fermat’s last theorem, 10 Hecke algebra generators over Z, 175 Hecke operator, 54, 225 Heilbronn matrices, 48, 132, 133, Krylov subspace, 114 lattices, 11 linear symmetric spaces, 245 Maeda’s conjecture, 28 Manin symbols, 44 and boundary space, 134 and cuspidal subspace, 134 modular symbols finite presentation, 44 new and old subspace of, 143 newform, 155 associated period map, 177 computing, 159 system of eigenvalues, 166 new modular symbols, 143 number field sieve, 69 old modular symbols, 143 partitions, 11 period mapping computation of, 185 Petersson inner product, 59, 160 Ramanujan graphs, 10 right action of GL2 (Q), 5 Serre’s conjecture, 11 Shimura-Taniyama conjecture, 37 valence formula, 17

Index